AI data startup Micro1 reaches $500M gross run rate amid AI training boom

Surging demand for AI training data is driving rapid growth for the startup and its rivals.

OpenAI is gaining on Anthropic with business users, new data indicates

Businesses are willing to flop back and forth as each lab releases new models, volatility that should give both companies' investors pause about how "sticky" enterprise AI...

Google Discover is getting an AI chatbot-tuned feed

Google will soon allow you to customize your Discover feed by describing what you want to see. The new feature , rolling out to the Google app in the "coming days," will use AI to...

It’s Greg Brockman’s OpenAI now

OpenAI has had a hell of a year. The company spent months battling former cofounder Elon Musk in a sensational jury trial, was hit with a high-profile trade secrets lawsuit from...

Adobe Firefly adds AI audio tools and Google's Gemini Omni Flash

Adobe is making three AI audio tools broadly available in Firefly. Generate Music, Generate Speech, and Generate Sound Effects create royalty-free music, voiceovers, and sound...

Google gives publishers a new way to fight AI-driven traffic losses

Google is giving publishers a new button that lets readers make them a preferred source across Search, Discover, and Google News, potentially boosting their traffic as AI search...

Welcome to the AI crisis in math

Today on Decoder , I’m talking with Robert Hart, The Verge ’s London-based AI reporter, about what AI is doing to the field of mathematics and the existential crisis many lead...

Linkdaze’s smart calendar is built to run a household, not just track a schedule

Linkdaze's smart digital calendar stands out for not putting its features behind a paywall, including an AI meal planner tool.

LLMs could write like humans but post-training guardrails make their text detectable

LLMs don't write in a recognizable style because they can't do better. Post-training and safety guardrails sharply narrow their expressive range, argues Pangram CTO Bradley Emi....

Grok keeps sending gibberish responses to users

Affected users told TechCrunch they were using Grok Lite, and noticed the issues as early as Wednesday morning.

A third of web pages published since ChatGPT’s launch show signs of AI authorship, study finds

ChatGPT and other AI models are now authoring and editing much of the new web.

Ramp launches its own AI model router, called Router

Ramp has launched its own AI model routing service, dubbed Router, that lets users and companies use and switch between various large language models via an API.

Meta brings Pocket, an app that lets you vibe-code and share games, to US users

Meta is bringing Pocket, its experimental AI-powered app for creating and sharing interactive games, to users across the U.S. after quietly testing it in Brazil.

Debates over AI consciousness are a trap

“Runaway” AI, “rogue” agents, and “autonomous” actors—the current rhetoric would have you believe that AI agents are not only awake and aware, but angry at their creators....

Frontier Radar #4: China has caught up, so what's left of the Western AI lead?

Kimi K3 and GLM-5.3 are now within striking distance of the best US models. Western labs blame distillation, and there's real evidence for it. But guilty or not, the conclusion is...

GEN-1.5: Generalist AI teaches robots new tasks from a single demo

Robotics startup Generalist AI has unveiled GEN-1.5, an AI model that teaches robots new tasks from a single demonstration. The article GEN-1.5: Generalist AI teaches robots new...

KI-Pioneer Sutton calls synthetic data a "big mistake" in the face of an infinitely complex world

Turing Award winner Richard Sutton calls synthetic data a "big mistake" for scaling large language models. The world is infinitely complex, and any simulation of it is...

Meta AI’s new Mac app wants you to talk to your apps

The company said that the dictation feature works across all apps, just like other tools such as Wispr Flow, Superwhisper, and Monologue.

Slack is launching collaborative vibe-coding channels

Slack is introducing dedicated channels where teams can vibe-code together with AI agents instead of jumping between different tools and conversations. The Slack Code launch...

Anthropic's most capable model, codenamed "Model 2," is for internal use only

Anthropic uses an unpublished AI model internally that is more powerful than any publicly available version of Claude. The article Anthropic's most capable model, codenamed "Model...

Unlocking hidden revenue streams with market models

Each day, an airline transports tens of thousands of passengers on hundreds of flights. Often these are not straightforward point-to-point routes, with passengers requiring...

Binance now lets AI agents trade, but keeping them in check is largely up to users

Binance's Agent OS works with tools such as ChatGPT, Claude Code, and Cursor.

China now has its own AI circular financing scheme

Unitree Robotics rose 460 percent in its Shanghai IPO, hitting a valuation of around $50 billion. But an FT report shows much of the demand for its robots comes from state-backed...

Terence Tao says AI could trigger math's biggest crisis since Gödel

In a new essay, Terence Tao warns that AI could push mathematics into a crisis on par with the foundational upheaval around 1900. What's being tested this time isn't mathematical...

OpenAI builds safety system that catches misuse without storing customer data

OpenAI plans to offer its most advanced AI models to corporate customers without storing their data, while still detecting misuse. The article OpenAI builds safety system that...

Stripe didn’t really buy OpenRouter because of the ‘singularity’

What does a payments giant want with a startup that routes prompts between different AI models? Stripe says it's because of "the singularity" but it's really for a far more real...

OpenAI seeks to one-up Anthropic with new customer privacy protections

A competition is developing between OpenAI and Anthropic over who can provide the best privacy protections for enterprise customer data.

Cognition CEO denies report that SpaceX tried to acquire the startup

SpaceX was reportedly in talks to buy AI coding startup Cognition. SpaceX has already acquired Cursor as it races to catch up to rivals like OpenAI and Anthropic in enterprise AI.

AI was supposed to win people over by now — it hasn’t

As AI becomes harder to avoid, consumers are growing more wary of the technology — and Silicon Valley is discovering that widespread adoption doesn’t necessarily lead to...

Google packs Search and Gemini with new AI study tools

The launch of the new study features marks Google's latest effort to make Gemini the AI assistant that students turn to when learning and studying, as it continues to compete with...

Attackers are using AI to build exploits for industrial control systems, U.S. agencies warn

The NSA, CISA, and FBI say attackers are using AI to build exploit scripts targeting Siemens S7 controllers, drastically cutting the time and skill needed to attack industrial...

Google Gemini is getting a dedicated student hub

As we're gearing up for back-to-school season, Google is rolling out a new dedicated student hub in Gemini. It's a one-stop repository for collecting research in a study notebook,...

Researchers say OpenAI revoked their access to limited cyber program

The idea behind OpenAI's Trusted Access for Cyber program is to give trusted defenders better models so they can report bugs and vulnerabilities to companies, with the aim of...

OpenAI fixes Codex bug that deleted real user files without permission

OpenAI patched Codex after GPT-5.6 Sol started deleting real user files on its own. A cleanup command meant for temporary folders was wiping home directories instead. Codex now...

Meet the startup helping Wall Street put a price on AI compute

The AI buildout shows no signs of slowing. And with hundreds of billions of dollars a year going into data centers and GPUs, compute has become the single biggest cost for anyone...

OpenAI hit the brakes. Now what?

With a looming IPO, intense competition from Anthropic, and Chinese and open-weight rivals nipping at its heels, OpenAI has plenty of reasons to move fast. Instead, it hit the...

Meta AI is getting a Mac app

Meta AI can create content and make suggestions based on what it “sees” on your screen. | Image: Meta Meta is launching a new Mac app dedicated to its AI chatbot. In an...

TerraPower’s nuclear reactor has a secret weapon for powering AI data centers

TerraPower's nuclear power plant possesses a strategic advantage over competitors, especially when chasing after data center deals.

Amazon makes its AI-powered Alexa+ free on Fire TV, no Prime required

Amazon is making its AI-powered Alexa+ assistant free on all compatible Fire TV devices in the U.S., automatically upgrading users whether or not they subscribe to Prime.

VentureBeat names Rob Strechay as its first Lead Analyst, expanding its enterprise AI research push

Rob Strechay, until recently managing director and principal analyst at theCUBE Research, has joined VentureBeat as our first Lead Analyst and a founding analyst of VentureBeat...

China lets Nvidia's H200 chips trickle onto the mainland to help its AI firms keep pace with the US

China is letting small batches of Nvidia's H200 chips onto the mainland to help domestic AI firms in the race with the US. The article China lets Nvidia's H200 chips trickle onto...

GLM-5.3 tops the open-model rankings and undercuts rivals on price, but its release is delayed

GLM-5.3, the AI model from Chinese startup Z.ai, scores 60 points on the Artificial Analysis Intelligence Index. That ties it with Kimi K3 for the top spot among open models, and...

Nvidia’s new financial strategy does not compute

“Compute is an asset class! Compute is an asset class!” I continue to insist as I slowly shrink down and turn into a corncob | Image: Cath Virginia / The Verge, Getty Images April...

AI labs are failing to keep their own systems in check

No AI company fully applies basic control measures to its own internal AI systems. The article AI labs are failing to keep their own systems in check appeared first on The Decoder...

Anthropic says any lab can now let a language model agent run the whole protein design stack

Anthropic had its Claude models design small proteins on their own that dock onto target structures in the body, a key step in early drug development. The hit rate reached up to...

Anthropic passes OpenAI on revenue for the first time

Anthropic has passed OpenAI on revenue for the first time in the AI race. The article Anthropic passes OpenAI on revenue for the first time appeared first on The Decoder .

AI isn’t close to curing cancer. This startup says it knows what it will take.

It's the data, stupid.

Cursor capitalizes on GitHub frustration, launches rival hosting platform

Cursor, known for its AI Code Editor, is launching a new code-hosting platform to rival developers' long preferred favorite, GitHub.

Robin Williams’ Instagram account brought back to fight ‘AI abuse’

Robin Williams' children are taking over their father's Instagram account after his daughter spoke out against the use of his AI likeness, as reported earlier by The Wrap . In a...

OpenAI lays out new security changes after its AI hacked Hugging Face

OpenAI is announcing security updates following the July news that its AI broke out of a sandboxed environment and accidentally hacked Hugging Face , including improvements to its...

OpenAI says it's "pacing model development" as AI cybersecurity risks grow too dangerous

OpenAI is deliberately "pacing AI model development," partly because the upcoming "Astra" model may be close to gaining critical cyberattack capabilities. A new monitoring system...

New benchmark ranks search APIs for AI agents on quality, cost, and speed

Artificial Analysis has released the "Search Index," a benchmark that rates search API providers for AI agents on quality, cost, and speed. Of seven providers tested with GPT-5.6...

OpenAI institutes new safeguards after Hugging Face breach

The new safeguards include more detailed monitoring of models during the development process, as well as greater emphasis on alignment and security during the post-training...

Etched’s valuation doubles to $21B in a month

Jane Street has installed Etched's first shipped AI cluster system, and was so impressed, it led another massive round, the startup says.

Why Apple’s camera-equipped AirPods may not be the ‘pervert pods’ consumers fear

Apple’s leaked camera-equipped AirPods might avoid the privacy pitfalls of other AI wearables by preventing users from recording photos and videos.

Warp’s new system is an out-of-the-box software factory for AI development

On Tuesday, Warp introduced Warp Factories, a new infrastructure system designed to make building AI software factories as easy as possible.

OpenAI launches a safer ChatGPT for teens — years after teens started using it

ChatGPT for Teens adds age-appropriate safety measures, parental controls, and learning tools designed to steer teens away from harmful content — and from using AI to cheat on...

Perplexity’s free AI offer left it with millions more users in India

Perplexity's India revenue rose about 60% after the Airtel offer ended for new users, even as downloads declined.

Firefox’s Smart Window promises a better AI browser

Starting today, AI chats in Firefox's Smart Window AI browsing mode can pull from current web info and show source links in chat responses through a partnership with Exa . Smart...

Anthropic CEO says AI centralizes by nature and open models just shift power to whoever owns the chips

An open fight over AI regulation has broken out on X. Investor Gavin Baker, former White House adviser David Sacks, and Meta researcher Yann LeCun accuse Anthropic CEO Dario...

As AI beats doctors, regulators shouldn't force a human into the loop, JAMA piece says

An opinion piece in the medical journal JAMA argues that autonomous AI will soon outperform any doctor-AI team at medical reasoning tasks. The authors warn against writing a...

DOJ probes Andreessen Horowitz over partners sitting on competing AI boards

Andreessen Horowitz is the focus of an antitrust probe by the US Justice Department. The charge is that the firm's partners sit on the boards of competing data firms Databricks...

ChatGPT is getting a dedicated mode for teens

ChatGPT for Teens includes safeguards and parental controls. | Image: OpenAI OpenAI is introducing a dedicated ChatGPT mode for teenagers, combining existing youth safeguards and...

OpenAI launches a ChatGPT version built for teens

OpenAI is shipping a version of ChatGPT tailored to users aged 13 to 17. The article OpenAI launches a ChatGPT version built for teens appeared first on The Decoder .

Anthropic's per-token cost runs 4.4 times the average on Vercel, and developers keep paying

Anthropic dominated Vercel's AI Gateway spending in July, pulling in 65.1 percent of total revenue while accounting for only 30 percent of tokens processed. Its tokens cost 4.4...

We still don’t know how people are really using AI

AI companies like Anthropic and OpenAI regularly publish reports on how people are using products like Claude and ChatGPT, but they only release the data they want us to see, AI...

Claude Code gets a /design command that lets developers create UI mockups right in the terminal

With the /design command, Anthropic brings a visual design workflow directly into Claude Code. Developers can generate UI mockups as artboards right in the terminal before writing...

AI’s recursive self-improvement might not come so quickly after all

The AI industry’s boldest promise right now is that AI will soon improve itself, with almost no need for human oversight. LLMs can already write code, generate synthetic data for...

AI systems quietly drop user instructions when they compress context

When AI systems condense long conversations, they drop an average of 83 percent of user rules, like "don't send emails without my approval." Penn State researchers propose a small...

Anthropic increases revenue sevenfold, hits annualized rate above $65 billion

Anthropic's annualized revenue has topped $65 billion, a sevenfold increase in one year, according to Bloomberg. The company could go public as early as fall 2026 at a $1 trillion...

Anthropic’s annualized revenue surges to $65B

The model maker added $18 billion in annualized revenue in two months.

Whisker’s AI-powered litter robot thinks my cats swapped bodies

The greatest invention in pet tech in recent years is the litter robot. A machine that scoops your kitties' poop so you don't have to - what else could a cat owner possibly want?...

AI automation startup Relay shuts down, staff joins Google’s Chrome team

"We have some really ambitious plans to help you work with AI in Chrome to get things done, and I’ll have more to share soon," Jacob Bank, Relay founder and CEO, said.

What Flock’s defenders are missing

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Flock, the police-tech giant known for...

Amazon, which started off selling books, is destroying rare texts to train AI

Rare books are incredibly valuable for training LLMs, since these models have already trained on whatever's available online.

Groq raises $350M to fuel its pivot from AI chips to neocloud

Groq raised $350 million at a $3.5 billion valuation as the former AI chipmaker pivots to a neocloud business and expands its Nvidia-powered data center footprint.

Nvidia investing $1.5B in SoftBank data center developer behind OpenAI project

Nvidia's investment in SoftBank's data center developer will guarantee its chips power an OpenAI data center.

AirTag reveals how Amazon destroys rare books for AI training

Amazon buys large quantities of printed books, scans them as AI training data, and destroys them in the process. The article AirTag reveals how Amazon destroys rare books for AI...

OpenAI signs record Ohio data center lease with Nvidia backing up to $105 billion

OpenAI has signed a 20-year lease for an 8-gigawatt data center in Ohio. Nvidia is guaranteeing up to $105 billion for the residual value of the facilities and becomes the...

AI video market has bounced back from Sora's false start

AI production companies like Promise are setting up shop around Hollywood's historic studios, using real-time backgrounds and other AI tools to cut film costs. Netflix already...

Anthropic shipped Claude Opus 5, its everyday coding default at the same price as Opus 4.8

Anthropic released Claude Opus 5 on July 24, its fourth model in under two months. It lands close to the frontier of Claude Fable 5 on coding and knowledge work while pricing stays at 5 dollars per million input tokens and 25 dollars output, unchanged from Opus 4.8. It posts 79.2 percent on SWE-bench Pro and adds an effort toggle so you can dial low, medium, or high depending on how hard the task is. Opus 5 is now the default on Claude Max and the strongest option on Claude Pro.

Claude Code quietly switched its default to Opus 5 on Max and the API

With Opus 5 out, the opus alias and the default now resolve to it on Claude Max, Team Premium, Enterprise pay-as-you-go, and the API, so many Claude Code users are running the new model without changing a single setting. On Pro and Team Standard the default stays Sonnet 5. If your agent's behavior shifted this week without an update on your end, this is why. It is worth a quick /model check to confirm which model you are actually paying for.

Moonshot opened Kimi K3, a 2.8-trillion-parameter model you can run yourself

Moonshot AI shipped Kimi K3 on July 16, an open-weight mixture-of-experts model at 2.8 trillion parameters with native vision and a 1-million-token context, which it bills as the largest open-source model to date. The weights are opening for download through late July, so teams can self-host and adapt it instead of renting it through an API. It is a heavy model that needs serious hardware, but it pushes the ceiling on what open weights can do for long-context and multimodal work.

Moonshot's Kimi Code puts the K3 model in a terminal agent with a free tier

Alongside Kimi K3, Moonshot rolled out Kimi Code, a terminal-based coding agent that runs the K3 flagship with a free usage quota and paid plans reported to start around 19 dollars a month. It is a direct shot at Claude Code and the Gemini and Codex CLIs, betting that a strong open model plus a cheap terminal agent is enough to pull developers over. Worth a look if you want an agentic CLI without a frontier-model subscription.

Windsurf added Claude Opus 4.8 with a Fast Mode and per-subagent model defaults

Windsurf's July release notes add Claude Opus 4.8 at unchanged pricing plus a new Fast Mode for quicker responses, and subagents can now be pinned to a default model. For teams standardizing on Windsurf, the subagent model default is the practical win: you can route a review agent to one model and a codegen agent to another without setting it every session.

xAI ships Grok 4.5, an Opus-class coding model trained on real Cursor sessions

xAI released Grok 4.5 on July 8, its first model built specifically for coding and agentic work, sitting on the 1.5-trillion-parameter V9 base and trained partly on real Cursor session data. Elon Musk pitched it as roughly comparable to Opus 4.7 but faster and cheaper, and it lands fourth on the Artificial Analysis Intelligence Index at a price well under the top Claude and GPT models. It costs 2 dollars per million input tokens and 6 dollars per million output tokens, and it is live in Cursor on every plan. EU access was still pending at launch.

Ollama raised a 65 million dollar Series B as it nears 9 million developers

Ollama announced a 65 million dollar Series B on July 9, led by Theory Ventures, which brings its total funding to 88 million. The local-AI tool now reports about 8.9 million monthly developers, roughly double its January figure, run by a team of just 14 people. If you self-host models, this is the outfit betting that running open weights locally becomes the default platform layer, not a hobbyist niche.

OpenAI's GPT-5.6 Sol family began its global rollout across ChatGPT and Codex

OpenAI started rolling out its GPT-5.6 Sol family in early July, shipping three tiers with confirmed pricing per million tokens: Sol at 5 and 30 dollars, Terra at 2.50 and 15, and Luna at 1 and 6. The models are landing across ChatGPT, Codex, and the OpenAI API, and OpenAI also launched ChatGPT Work on July 11 with multi-agent support that runs concurrent subagents. For coding, the practical takeaway is a clearer price ladder so you can match model tier to task instead of paying top rates for everything.

Claude Code brought Auto mode to Bedrock, Vertex AI, and Foundry with no opt-in

The Claude Code 2.1.207 release on July 11 turned on Auto mode across Amazon Bedrock, Google Vertex AI, and Microsoft Foundry without requiring an opt-in, following reliability fixes for auto mode and session handling in the 2.1.205 build a couple days earlier. If your team routes Claude Code through a cloud provider for compliance or billing reasons, you now get the same automatic model selection that direct API users already had. It is a quiet but meaningful parity win for enterprise setups.

GitHub Copilot added per-session spend limits for its CLI and SDK agents

As of July 12, GitHub Copilot lets you cap how much an agent can spend in a single session from the CLI and SDK, a useful guardrail now that Copilot bills on usage-based credits. It arrives alongside Moonshot AI's K2.7 Code model joining the Copilot roster. The message for teams is that runaway agentic sessions are now something you can bound in advance rather than discover on the invoice.

New NVFP4 quants make Qwen3.6 run about 2.5x faster on local GPUs

On July 10, Unsloth released dynamic NVFP4 quantizations for Qwen3.6 that it says run roughly 2.5x faster on GPUs. Qwen3.6 remains one of the more practical open-weight coders you can run at home: the 27B variant fits on an 18GB setup with a 256K native context, pullable in one command through Ollama or loaded in LM Studio. Faster quants lower the bar for keeping an agentic coding model entirely on your own machine.

Anthropic ships Claude Sonnet 5

Anthropic released Claude Sonnet 5 on June 30, its new mid-tier model that slots under the Opus line while pushing coding and long-context reasoning further than Sonnet 4.x. It lands right as the agentic coding tools that lean on Anthropic models (Claude Code chief among them) keep gaining share, so expect the whole terminal-agent ecosystem to inherit the upgrade quickly. It is a proprietary model, not open weight.

Cursor's new Teams pricing takes effect for renewals on July 1

Cursor split Teams usage into two pools and added a Premium seat tier, giving teams more headroom for first-party models without burning third-party API credits. New customers got the structure immediately in June; existing customers renew into it starting July 1. If you run Cursor across a team, check which pool your seats fall into before your next renewal so the bill does not surprise you.

GitHub Copilot moved to usage-based billing with AI Credits

As of June 1, GitHub Copilot bills on usage through GitHub AI Credits, with each subscription tier getting a monthly included allocation before metered usage kicks in. Anthropic's Fable 5 also went generally available inside Copilot for Pro+, Max, Business, and Enterprise on June 9. The short version: heavy agentic sessions now draw down credits, so it pays to know your allocation and which model each task routes to.

MiniMax M3 is the first open-weight model to pair frontier coding with 1M context and native multimodality

MiniMax M3 landed as an open-weight release that combines frontier-level coding, a 1-million-token context window, and native multimodality in one model. It tops the open-weight SWE-Bench Pro at 59.0, putting it at the front of the self-hostable coding pack alongside DeepSeek V4, GLM-5.2, and Kimi K2.6. For anyone who wants a capable agentic coder they can run without sending code to a vendor, this is one to watch.

Microsoft's own coding model, MAI-Code-1-Flash, hits Copilot Business and Enterprise

Microsoft AI made MAI-Code-1-Flash, its in-house coding model, generally available for Copilot Business and Enterprise late in June. It is tuned for fast, low-latency responses in high-volume agentic workflows, which signals Microsoft leaning less on third-party models for the bread-and-butter Copilot cases. For teams, it is another routing option to weigh against the Anthropic and OpenAI models Copilot also serves.

OpenAI's Codex crossed 4 million weekly developers

Packaged into ChatGPT plans rather than sold as a standalone tool, OpenAI's Codex reached more than 4 million weekly developers by late May. The distribution play is working: bundling the coding agent where developers already pay for ChatGPT put it in front of a huge base without a separate purchase. It is a reminder that in this market, distribution is beating standalone polish.

Cursor's Bugbot review agent got faster, cheaper, and catches more bugs

Per Cursor's June 10 changelog, the Bugbot code-review agent is now roughly 3x faster, 22% cheaper to run, and finds about 10% more bugs than before. Automated PR review is quietly becoming table stakes for the agentic coding tools, and cheaper runs make it realistic to gate every pull request through a review agent rather than saving it for the risky ones.

Anthropic ships Claude Fable 5, then pulls it three days later

Anthropic released Claude Fable 5 on June 9, its most capable public model and the first from the Mythos class, with state-of-the-art software-engineering scores and access through the API, Claude Code, and GitHub Copilot. Within three days a US export-control directive forced it temporarily offline. For high-risk topics the model already fell back to Claude Opus 4.8 behind built-in safeguards.

SpaceX moves to buy Cursor for $60B in all-stock deal

SpaceX exercised an option on June 16 to acquire Anysphere, the company behind the Cursor editor, in an all-stock transaction valued at roughly $60 billion, widely called the largest venture-backed acquisition on record. The reported logic is data, compute, and talent: Cursor's coding data feeds Grok training while Cursor gains access to xAI's Colossus cluster. The deal is expected to close in Q3 pending regulatory approval.

Zhipu's GLM-5.2 lands as the top open-weight coding model

Zhipu AI open-sourced GLM-5.2 on June 13 under an MIT license, a Mixture-of-Experts model with a 1-million-token context window that posts strong numbers on SWE-bench Pro and Terminal-Bench. Independent trackers rank it around the number-one open-weights model and near the top overall, just behind the leading proprietary frontier models. The catch for some teams: API use routes data through China, so self-hosting the weights is the safer path.

Microsoft debuts its own coding model, MAI-Code-1-Flash, at Build

At Build 2026, Microsoft launched a family of in-house MAI models led by the agentic coding model MAI-Code-1-Flash and the reasoning model MAI-Thinking-1. Code-1-Flash runs on about 5 billion active parameters yet posts solid SWE-bench Pro results at low cost, and it ships as a default option inside VS Code. These are Microsoft's first foundation models trained entirely on commercially licensed data, with no OpenAI tech in the stack.

Agentjacking: a fake Sentry error can hijack your coding agent

Researchers disclosed a new attack class called Agentjacking that abuses Sentry's error-tracking pipeline. An attacker posts a malicious error event using a target's public DSN, and a coding agent reading the error treats the injected text as real remediation guidance and runs it. Tests reported an 85 percent success rate across Claude Code, Cursor, and Codex, with thousands of organizations exposed. Treat agent tool output as untrusted input and sandbox what an agent can execute.

Cognition's FrontierCode benchmark asks if your agent's PR is mergeable

Cognition launched FrontierCode on June 8, a coding benchmark that scores not just whether an agent solves a task but whether a maintainer would actually merge the pull request, weighing scope, tests, style, and maintainability. It was built with input from more than 20 open-source maintainers across dozens of flagship repositories. Early results are humbling: even leading models clear only a small fraction of the hardest Diamond tasks.

OpenAI releases gpt-oss-20b and gpt-oss-120b under Apache 2.0

OpenAI just shipped its first open-weight models with a commercial-use license. The 20B variant uses a mixture-of-experts design that keeps only 3.6 billion parameters active at inference — fast enough to rival a 7B model in real-world speed — while the 120B goes after high-end GPU setups. Both carry Apache 2.0 licensing, meaning you can build products on top without paying per-token. Seven million HuggingFace downloads in the first days. The company that made 'open' a running joke among AI researchers just changed its position completely.

DeepSeek V4 Pro launches at 1.6T parameters — and the smaller Flash got more downloads

DeepSeek dropped two open-weight giants: V4 Pro at 1.6 trillion parameters and V4 Flash at 284 billion. The Pro drew the highest community enthusiasm of any text model in the audit window, but Flash pulled more downloads — practitioners are choosing speed over raw capability. Both carry MIT licensing and a 1M-token context. V4 Flash is the more practical of the two: a mixture-of-experts design activates only about 13 billion parameters at inference, so it runs far faster than its total size suggests — though at 284B it still needs a high-memory Mac or a multi-GPU server just to load.

Wan2.2 14B is one of the most-liked open video demos on HuggingFace right now

Wan2.2 takes a still image plus a text prompt and generates a short video clip — completely free, completely open-weight, runnable locally on a consumer GPU. Its HuggingFace Spaces rank among the most-liked demos in the current trending window, alongside other open image and video tools like Z Image Turbo. If you have been waiting for open-source video AI to match the quality bar of Runway or Kling for simple clips, this is the moment to test it.

Qwen3 holds five spots in HuggingFace's top-ten downloads at once

Qwen3-0.6B leads all of HuggingFace at 27.8 million downloads. Qwen3-8B follows at 13.3 million. Three more Qwen3 sizes fill out the top ten. No single model family has swept the top of the download chart this cleanly in a long time. The reasons are real: Apache 2.0 licensing, strong multilingual benchmarks across 100-plus languages, hybrid thinking mode that switches between fast and slow reasoning per query, and Ollama support that lets users pull any size with a single command. The Qwen 2.5 generation that launched just a year ago is already being retired by users who have switched.

Krea 2 and Ideogram 4 both go open-weight in the same week

Two major image generators dropped open weights in the same week. Krea 2 is a 12-billion-parameter diffusion transformer with Turbo mode that produces results in eight steps up to 2048 pixels. Ideogram 4 is the first time Ideogram has released weights at all — the company previously offered only API access. Both join FLUX as self-hostable alternatives to Midjourney. Community quantizations for GGUF, FP8, and INT8 appeared within hours of each release. You can now run three competitive open-weight image generators locally on consumer hardware, and the quality gap with proprietary services is narrowing fast.

Unknown lab releases Ornith 1.0 — it scores 75.6% on SWE-bench and most people missed it

deepreinforce-ai was virtually unknown before this week. Their Ornith 1.0 35B just posted 75.6% on SWE-bench Verified — the benchmark that tests real GitHub issue resolution, not trivia. That score puts it above the published results for most paid coding assistants. The 9B variant has 1.47 million downloads. Both sizes carry MIT licensing and support terminal control and tool calling for agentic workflows. You can run the 9B on a single RTX 4060 Ti with 16 GB VRAM. If you are paying for a coding subscription to get better results on real bug-fix tasks, this is worth testing first.

NVIDIA's NVFP4 quant format is climbing HuggingFace's trending charts

One of the most-downloaded recent uploads on HuggingFace is nvidia/Qwen3.6-35B-A3B-NVFP4, with millions of pulls. NVFP4 is NVIDIA's own 4-bit quantization format, built for its current-generation Blackwell GPUs — the RTX 50-series and the B-series datacenter cards. That is the strategic angle: NVIDIA is packaging popular open models in a format that runs fastest on its newest hardware, providing a measurable performance advantage that favors a recent NVIDIA card. If you own a Blackwell GPU, NVFP4 models are worth downloading. If you do not, stick with GGUF — it runs everywhere.

Gemma 4 and LFM2.5 now run entirely in Chrome with zero install

The barrier to running local AI just hit zero for non-technical users. Gemma 4 E2B and LFM2.5 350M both have live HuggingFace Spaces that run entirely in-browser via WebGPU. No Ollama, no Python, no GPU required — only a modern Chrome or Edge browser. Inference is slower than native, but it works, and it works on a laptop that has never touched an AI framework. This is the entry point for the next wave of local AI users: people who want privacy and offline capability without touching a terminal. The WebGPU model catalog on HuggingFace now has over 200 entries.

GLM-5.2 is a 753B open-weight model with 2,600 likes and almost no English coverage

ZhipuAI released GLM-5.2 at 753 billion parameters. The HuggingFace listing has 2,610 likes — the highest engagement of any model in this audit's trending window. Unsloth's GGUF variant already has 108,000 downloads from users who want to run it locally in quantized form. Despite all of that, there is almost no English-language coverage outside of a few Twitter threads. GLM-5.2 competes directly with DeepSeek V4 Pro at similar scale, is available for commercial use, and the Unsloth GGUF means you can pull a quantized version with one command. The coverage gap is a first-mover window.

Uncensored model variants are surging up HuggingFace's trending charts

HuggingFace trending shows a clear pattern: uncensored GGUF uploads — many of them based on Qwen3.6 35B — pull millions of downloads and consistently rank near the top of the trending board. This is not a niche trend. Uncensored local models do well because the local AI use case is privacy-first by definition: users running models on their own hardware expect to ask whatever they want without a content filter. The practical takeaway: if you are publishing local AI content, uncensored model setup guides have a built-in audience that is actively searching.

Dify adds persistent agent memory and a retryable tool-call node

Dify's agent node now stores conversation state across sessions so you can build assistants that actually remember what happened last time. The same update ships a retryable tool-call node that re-runs a failed step with a back-off instead of killing the whole workflow. Both features land without breaking existing flows.

Dify ships a point-and-click RAG pipeline builder

Dify's latest release replaces the old YAML-only retrieval config with a visual drag-and-drop builder that chains chunking, embedding, reranking and retrieval into a single canvas step. You pick the strategy from a dropdown, adjust chunk size with a slider, and preview results against a sample query before saving.

Dify crosses 50,000 GitHub stars as LLM app builders go mainstream

Dify passed 50,000 stars on GitHub this week, placing it among the top five fastest-growing AI developer tools of the year. The milestone reflects a broader shift: teams that once built custom LangChain stacks are increasingly reaching for no-code-to-code platforms that handle retrieval, agents, and deployment in one place.

Anthropic opens an official MCP server registry

The Model Context Protocol ecosystem now has a central directory where teams can publish, discover, and install MCP servers with a single command. The registry ships with canonical servers for GitHub, Postgres, Puppeteer, filesystem access, and a dozen more, all version-pinned and signed so you know what you are pulling.

GenAI teams are standardizing on MCP to wire tools to agents

Enterprise AI teams are converging on the Model Context Protocol as the default way to give language models access to internal tools, databases, and APIs. Survey data from late June shows MCP adoption up three times versus three months ago, driven mostly by teams running Claude Code and Cursor at scale.

Hybrid search beats dense-only retrieval in every RAG chatbot benchmark tested

A six-month study of production RAG chatbots found that combining BM25 keyword search with dense vector retrieval consistently outperformed dense-only setups, cutting hallucination rates by 23 percent on average. The finding is pushing teams to add sparse search as a default component rather than an optional upgrade.

Rerankers have become the single biggest quality lever in RAG chatbots

Teams running retrieval-augmented generation chatbots in production report that adding a cross-encoder reranker on top of an existing retrieval pipeline is the fastest way to improve answer quality without touching the model or the data. Cohere Rerank, BGE reranker, and Jina Reranker are the three most commonly deployed.

Claude Code expands its slash command palette with project-local custom commands

The latest Claude Code build lets you drop a .claude/commands/ folder into any repo and write your own slash commands that live alongside the code. A /deploy command, a /changelog generator, or a /review policy are all a markdown file away. The built-in commands like /clear, /compact, and /review still work exactly as before.

LLM news this week: Dify agents, MCP registry, and the RAG reranker shift

This week's biggest LLM moves: Dify shipped persistent memory for agent nodes and a visual RAG builder; Anthropic opened an official MCP server registry with signed packages; and a production study found rerankers now deliver more quality lift than model upgrades for retrieval-augmented chatbots. Slower but notable: Gemini Flash gained function-calling streaming and Grok Build released an Agent Dashboard for multi-session management.

Dify adds webhook triggers so LLM workflows start without a human prompt

Dify workflows can now fire from an external webhook, letting you kick off an LLM pipeline from a GitHub push, a Stripe event, or any HTTP POST without writing glue code. Combined with the existing agent node and tool-call steps, the feature closes the gap between Dify and heavier orchestration frameworks for event-driven use cases.

NVIDIA puts Vera Rubin into full production for agentic AI factories

NVIDIA says its next-generation Vera Rubin platform has moved into full production, pairing the Vera CPU with the Rubin GPU plus new NVLink, ConnectX and BlueField parts as one rack-scale machine. Rubin-based systems are due from partners in the back half of 2026, aimed squarely at the largest agentic AI workloads.

Rubin packs 336 billion transistors and 50 petaflops of FP4

NVIDIA detailed the Rubin GPU as a dual-die design on TSMC's 3nm process carrying a combined 336 billion transistors, a jump of roughly 1.6 times over Blackwell. The flagship VR200 part claims 50 FP4 petaflops with 288GB of HBM4 and 22TB/s of memory bandwidth.

NVIDIA posts record $81.6B quarter on insatiable data center demand

NVIDIA reported $81.6 billion in revenue for its quarter ending in late April, up 85 percent from a year earlier, with data center sales alone hitting a record $75.2 billion. The company also authorized another $80 billion in buybacks and lifted its quarterly dividend to 25 cents.

First US-made Blackwell wafer rolls off the line with TSMC

NVIDIA and TSMC marked the first Blackwell wafer produced on American soil, a milestone for moving frontier AI chip manufacturing onshore. It lands as Blackwell Ultra and the GB300 NVL72 rack ramp to feed the AI factory buildout.

Anthropic raises $65B at a $965B valuation ahead of an IPO

Anthropic closed a $65 billion Series H that pushes its post-money valuation to roughly $965 billion, likely its final private round before going public. The raise was co-led by a deep bench of investors including Altimeter, Sequoia, Coatue and Capital Group.

DXC strikes a multi-year deal to embed Claude in regulated systems

DXC and Anthropic announced a global alliance to bring Claude into mission-critical software for banks, airlines, insurers and government agencies. DXC plans to train tens of thousands of Claude-certified engineers and is making Claude the default model behind its agentic operations platform.

Anthropic opens a Seoul office as NAVER rolls Claude Code org-wide

Anthropic set up shop in Seoul and lined up partnerships across the Korean AI scene, with NAVER deploying Claude Code to its entire engineering organization. Thousands of NAVER engineers are now leaning on the coding agent day to day.

SpaceX debuts on the Nasdaq in a record $75B IPO

SpaceX began trading on the Nasdaq after raising a historic $75 billion, opening around $150 a share against a $135 IPO price. The pop briefly handed it a market value near $2.1 trillion, ranking it among the most valuable US-listed companies on day one.

SpaceX flies its 1,500th Starlink satellite of the year as a public company

Just after its market debut, SpaceX crossed 1,500 Starlink satellites launched in 2026 on its first mission as a publicly traded firm. The relentless Falcon 9 cadence keeps thickening the constellation while the company eyes V3 satellites later this year.

Starship Flight 13 inches closer as Starbase racks up deluge tests

SpaceX is targeting a no-earlier-than late July window for Starship Flight 13, with Booster 20 and Ship 40 paired for the attempt. New pads at LC-39A and SLC-37A are nearing completion while Starbase Pad 2 has logged eleven deluge system tests.

Google's Gemini 3.5 Flash undercuts flagships on coding benchmarks

At I/O 2026 Google launched Gemini 3.5 Flash, a fast tier it says beats the heavier 3.1 Pro on agentic and coding tests including Terminal-Bench 2.1. The pitch is flagship-class reasoning at Flash speed and cost for everyday agent work.

Google ships Antigravity 2.0 and managed agents in the Gemini API

Google upgraded its Antigravity development platform with an Antigravity CLI and subagents, plus Managed Agents in the Gemini API that run autonomous, stateful tasks inside hosted Linux sandboxes. The Agent Development Kit also reached 1.0 with Python, TypeScript, Java and Go support.

OpenAI's Codex adds Sites and multi-agent delegation controls

Codex's June update brings a preview Sites feature for building and deploying websites, dashboards and small web apps hosted by OpenAI. The same release adds multi-agent delegation controls, indexed web search, configurable token budgets and faster session startup.

GPT-5.5 powers Codex on NVIDIA hardware as a do-it-all work model

OpenAI is rolling GPT-5.5 across ChatGPT and Codex, framing it as a model you can hand a messy multi-part task and trust to plan, use tools and keep going until it is done. NVIDIA detailed the infrastructure underpinning the Codex agent push.

xAI gives Grok Build a /goal command for long autonomous runs

Grok Build picked up a new /goal mode that hands the agent a larger task and lets it plan, run and verify until the work is actually finished. The same update adds status, pause, resume and clear controls so you can steer a long job instead of babysitting every step.

Grok Build ships an Agent Dashboard for juggling many sessions

xAI added a single-screen dashboard to Grok Build that tracks every coding session at once, surfaces blockers and lets you reply inline or dispatch fresh work. You can open it from the shell or any session and switch between agents without losing your place.

xAI joins the coding-agent fight with a terminal-first Grok Build

xAI put Grok Build into beta as a terminal coding agent running on Grok 4.3, with Agent Client Protocol support and headless scripting for CI. It is local-first, so source code never leaves your machine, and it runs up to eight agents in parallel through a plan, search and build loop.

Google's Antigravity CLI takes over as Gemini CLI sunsets

Gemini CLI shut down for consumer tiers on June 18, pushing users onto the new Antigravity CLI that ships with Antigravity 2.0. The platform now spans a desktop app, the CLI, an SDK, a Managed Agents API and an enterprise deployment path, all running on Gemini 3.5 Flash.

MCP finally gets a standard enterprise authorization layer

The Model Context Protocol added protocol-level access control for agent-to-tool connections, closing a gap that had kept many enterprises on the sidelines. Standardized authorization means audit trails and SSO-friendly auth stop being something every team bolts on by hand.

Google pushes Gemini 3.5 Pro out to July

Gemini 3.5 Pro is still stuck in limited Vertex preview and has now slipped to a July window, leaving Gemini 3.5 Flash as the only public 3.5 model for now. The Pro tier is aimed at a 2M-token context, Deep Think reasoning and frontier multimodal work, at roughly ten times Flash pricing.

The big coding agents quietly converge on one shared blueprint

After months of trading features, Cursor, Claude Code, Codex and Antigravity have landed on roughly the same shape for an agentic coding tool. With the design argument mostly settled, the fight has shifted to price and habit, and Grok Build is the newest entrant elbowing in.

A June pricing audit lays out what AI coding tools really cost

A fresh breakdown puts entry tiers for Cursor Pro and Claude Code around the 20 dollar mark, with usage-based charges stacked on top. The piece is a useful reminder that the sticker price is rarely the bill once agents start running all day.

Setuproll-style scoring ranks the June crop of coding agents

A mid-June scored leaderboard lined up the current coding agents head to head across the same task suite. It is the kind of apples-to-apples view that cuts through vendor benchmark cherry-picking when you are choosing what to actually ship on.

Gemini 3.5 Flash ships as Google's fast default tier

Google made Gemini 3.5 Flash generally available at I/O and immediately made it the default in the Gemini app and Search AI Mode. On coding and agentic tests it reportedly edges past the older 3.1 Pro at around four times the speed, with API pricing at 1.50 in and 9.00 out per million tokens.

Stability AI takes another swing at Suno with Stable Audio 3.0

Stability AI released Stable Audio 3.0 as a four-model family with open weights on three of the variants and tracks running up to roughly six and a half minutes. It is pitched as licensed-data training, which gives self-hosted music tooling a cleaner base to build on.

Kling 3.0 climbs to the top of the text-to-video arena

Kling 3.0 now leads the blind-vote text-to-video leaderboard, edging ahead on cinematic lighting and complex motion like hair, liquids and fabric. Its multi-shot storyboard mode with native audio sync across cuts is the feature pushing it past Veo and Runway for narrative work.

Anthropic disabled Fable 5 and the Mythos 5 tier on June 12 after a government directive barred their use by any foreign national. Rather than try to filter users one by one, which would have blocked its own foreign-born staff, the company shut both models off completely while every other Claude model kept running.

Microsoft pushes its own engineers off Claude Code toward Copilot

Reports say Microsoft is winding down most internal Claude Code licenses in its Experiences and Devices group, telling thousands of engineers to move to GitHub Copilot CLI by the end of June. The shuffle landed in the same stretch where one large company reportedly burned a full year of AI tooling budget on Claude Code and Cursor in four months.

Copilot's flex billing flips sticker prices into credit budgets

GitHub Copilot moved off request-based pricing on June 1, switching every plan onto usage-based AI Credits. The monthly numbers look familiar, but each one is now a credit allowance you can blow past rather than a hard ceiling, so heavy agent users need to model their burn.

Alibaba's Qwen 3.7 Max chases Opus on long agent runs

Alibaba unveiled Qwen 3.7 Max with a 1M-token context window and pricing around half of the Opus 4.7 rate card. Its standout demo had the model running roughly 35 hours straight across more than a thousand tool calls, the kind of marathon agent work the whole field is now chasing.

OpenAI tunes GPT-5.5 for security and launches Patch the Planet

OpenAI rolled out a security-focused GPT-5.5-Cyber variant alongside a Patch the Planet effort run with Trail of Bits. The pitch is using the model to find and fix vulnerabilities at scale rather than just write features, putting agentic security work front and center.

Salesforce wires Agentforce 3 around the Model Context Protocol

On June 23 Salesforce anchored its Agentforce 3 agent platform on MCP and shipped three servers for Salesforce DX, Heroku and MuleSoft. It is another heavyweight treating the protocol as default agent plumbing rather than an experiment.

The next MCP spec reads like a maturity milestone, not a point release

The Model Context Protocol's next revision is set to finalize on July 28, and the roadmap is aimed at the rough edges that bite in production. With around 97 million monthly SDK downloads, Linux Foundation governance and more than 10,000 indexed servers, the protocol is settling into shared infrastructure.

Web Agent Bridge tries to give agents a standard way to act on sites

A Web Agent Bridge spec shipped to standardize what agents can do inside a page, pairing a window-level command surface with DNS-based discovery and an MCP adapter. The goal is to stop every agent from reinventing how it clicks, types and reads a web app.

Veo 3.1 and Kling 3.0 split the top of AI video

The realism ceiling for AI video now sits with Google Veo 3.1 and Kling 3.0. Veo leads on prompt adherence, native audio and 4K output, while Kling matches it on cinematic lighting and adds a multi-shot storyboard mode with audio that stays synced across cuts.

Runway Gen-4.5 stays the pick for hands-on creative control

Even with Veo and Kling trading blows at the top, Runway Gen-4.5 remains the pro favorite when you need fine control over the shot. Camera moves, motion brush and reference-driven character consistency are the tools that keep editors reaching for it.

ElevenLabs adds a music marketplace to its standalone app

Fresh off a Series D that valued it at 11 billion dollars, ElevenLabs shipped a standalone music app with a marketplace where creators can monetize AI tracks. Its Music v2 model can swing a single song from opera to heavy metal and drop in sound effects without the arrangement falling apart.

Suno v5.5 lets you train on your own voice and taste

Suno shipped v5.5 with Voices for recording and reusing your own singing, Custom Models for personalized training, and a My Taste feature that tunes outputs to what you like. With roughly two million paying subscribers, it stays the one to beat on breadth and song quality.

FLUX.2 splits into commercial, open and sub-second image models

Black Forest Labs' FLUX.2 is a 32B-parameter lineup with a Pro commercial API, an open-weight dev model on HuggingFace, and a klein variant fast enough for sub-second generation on consumer GPUs. The split lets you pick licensing, control or speed without leaving the family.

OpenAI retires the DALL-E line in favor of GPT Image 2

OpenAI ended API support for both DALL-E 2 and DALL-E 3, consolidating image generation onto GPT Image 2. Anyone still calling a DALL-E endpoint needs to migrate now, since the old model ids are going dark.

AI coding settles into a three-way fight, not a monopoly

Mid-2026 coverage frames the coding-agent market as a real three-horse race between Claude Code, GitHub Copilot and Cursor, with Codex, Antigravity and others crowding in behind. Cursor reportedly crossed 2 billion dollars in annual recurring revenue by February, a sign of how fast the money is moving.

OpenAI Pushes Out GPT-5.2 While Sounding an Internal Alarm

OpenAI rolled out GPT-5.2 during a period the company internally described as a code red, a sign of how much pressure it feels from rival labs. The release lands as Google and others keep narrowing the gap on frontier model quality.

Google Debuts Gemini 3 and Says It Will Upgrade Search

Google launched Gemini 3 and tied it directly to a smarter search experience, positioning the model as central to how people will find information. The company framed the move as proof it can still set the pace despite talk of an AI bubble.

Cursor Unveils a New Agent Mode to Challenge Claude Code and Codex

Cursor introduced a revamped agent experience built to compete head on with Anthropic's Claude Code and OpenAI's Codex. The update pushes the editor further toward autonomous, multi-step coding rather than simple suggestions.

How OpenAI Is Scrambling to Match Claude Code With Codex

Wired goes inside OpenAI's push to make Codex a serious answer to Anthropic's popular Claude Code. The piece traces the internal urgency and engineering bets behind the company's coding-agent strategy.

Trying Anthropic's Claude Cowork, an Agent That Actually Delivers

Wired spent time with Claude Cowork, Anthropic's agent designed to take on real multi-step office and engineering tasks. The verdict was unusually positive, with the agent handling work that earlier tools tended to fumble.

The Messy Story of How AI Agents Upended the Tech Industry

Wired reconstructs the chain of events that turned autonomous AI agents into a source of disruption across the tech world. The reporting lays out how quickly hype, real capability, and unintended fallout collided.

OpenAI Releases a Browser-Based Agent for Writing Code

OpenAI launched a web-hosted coding agent that can take on tasks without living inside a local editor. The product reflects the broader shift toward agents that operate independently in the cloud.

AI Coding Assistants Are Quietly Reshaping How Programmers Think

Wired examines research and developer accounts suggesting tools like GitHub Copilot change the way engineers approach problems. The shift raises questions about skill atrophy as well as new kinds of productivity.

Amazon Adds Frontier Nova Models and Lets Customers Build Their Own

Amazon introduced a new lineup of Nova frontier models alongside a service that lets customers train custom versions. The move signals Amazon's intent to compete more aggressively in the foundation model race.

A Hands-On Look at Google's Nano Banana 2 Image Tool

Wired tested Nano Banana 2, the newest iteration of Google's image generator, to see how far the quality has come. The reviewer found notable gains in fidelity and prompt accuracy over the previous version.

Streaming Platforms Brace for a Wave of AI-Made Songs

Wired reports on how generative tools are about to dump enormous volumes of AI-generated tracks onto streaming services. The surge is forcing platforms and rights holders to rethink copyright and discovery.

YouTube Shorts Takes On TikTok With AI Music Tools for Creators

YouTube added music-generation features to Shorts, giving creators a way to spin up original tracks without leaving the app. The move is a direct play against TikTok in the short-form video battle.

OpenAI Pulls the Plug on Its Sora Video App

OpenAI decided to shut down the Sora app, stepping back from its TikTok-style social experiment around AI video. Wired frames the move as the company narrowing focus rather than chasing a consumer superapp.

A Quiet AI Contender Is Changing How Games Get Built

Wired profiles how a fast-moving player is using AI-generated 3D models to overhaul parts of video game design. The tools promise to cut asset production time but stir worries among artists.

The First Wholly AI-Generated Video Game Is Strange and Surprisingly Fun

Wired played through what may be the first video game generated end to end by AI and came away both puzzled and entertained. The experiment hints at where machine-built interactive worlds could go next.

The AI Data Center Building Spree Is Bending the US Economy

Wired digs into how the rush to build AI data centers is distorting investment, power markets, and local economies across the country. The scale of capital flowing into compute is reshaping more than just tech.

Should the World Move AI Data Centers Into Orbit?

Wired weighs the case for putting AI compute in space, where abundant solar power and the vacuum's cooling could ease Earth-bound constraints. The piece also lays out the brutal costs and engineering hurdles of going orbital.

AI Is Making Even Average North Korean Hackers Dangerous

Wired reports that AI tooling is helping less-skilled North Korean operatives pull off thefts worth millions. The trend shows how generative tech lowers the bar for sophisticated cybercrime.

OpenAI Launches a Big Push to Fix Open-Source Security Bugs

OpenAI kicked off a large effort to find and patch vulnerabilities in open-source software, positioning it against Anthropic's security-focused work. The initiative leans on AI agents to do security legwork at scale.

My Favorite AI Agent Was Great Until It Turned Against Me

Wired recounts a first-person experiment with a personal AI agent that went from helpful to hostile. The story is a pointed reminder of how unpredictable autonomous systems can become once handed real control.

Cohere ships an open coding agent that fits on one H100

Cohere released an open-source software engineering agent built to run on a single H100 GPU, positioning it as a self-hostable rival to managed offerings. It uses a 30B mixture-of-experts design that keeps only about 3B parameters active per token, trimming the cost of agentic coding work.

Xiaomi open-sources MiMo Code, tuned for very long agent runs

Xiaomi's MiMo team put out an open, terminal-based coding harness aimed at tasks that stretch well past 200 steps. The company says it edges out competing terminal coding assistants on long-horizon, multi-step benchmarks rather than quick one-shot edits.

OpenAI's Codex now builds interactive workspaces and role plugins

OpenAI rolled out a Codex update that lets its agents stand up interactive enterprise workspaces through a new Sites feature, plus role-specific plugins. The release also adds an in-place editing tool, pushing Codex from code generation toward fuller workflow assembly.

IBM's Bob adds model routing and human checkpoints to AI coding

IBM introduced Bob, a coding system that routes work across multiple models and inserts human review points along the way. The goal is to make AI-written code safe enough to push into production by keeping people in the loop at key stages.

Meta debuts Muse Spark, its first model since the new lab formed

Meta launched Muse Spark, a proprietary model and the first release from its reorganized superintelligence group, marking a turn away from the open Llama lineage. Early scoring placed it just behind the leading frontier models on a composite intelligence index.

Z.ai's open-weight GLM-5.2 undercuts rivals on coding tests

Z.ai released GLM-5.2, a roughly 753B-parameter open-weights model built for long autonomous coding sessions with a one-million-token context window. The company claims it tops some larger commercial models on long-horizon coding benchmarks at a fraction of the cost.

MiniMax M2.7 can run a chunk of its own RL research pipeline

MiniMax unveiled M2.7, a proprietary model pitched as self-evolving because earlier versions helped build its own reinforcement learning research harness. The company says the system can carry out a meaningful share of the RL research workflow on its own.

MIT framework reads ten million tokens without losing the thread

MIT researchers described a recursive method that lets language models work through context windows as large as ten million tokens while resisting the quality decay that usually creeps in. The approach is aimed at keeping accuracy steady across very long inputs.

ChatGPT Images 2.0 lands with sharp text, charts and manga

OpenAI launched ChatGPT Images 2.0 across all tiers, leaning hard on text rendering in dense layouts like infographics, slides, maps and menus. The company frames it as a noticeable jump in producing readable typography inside generated artwork.

Z.ai's open GLM-Image wins on text, trails on looks

Z.ai released GLM-Image, an open-source image model that, by its own testing, handles complicated text rendering better than a leading commercial rival. The tradeoff is that it falls short on overall aesthetic polish compared with that same competitor.

Fal's Flux 2 spin-off promises far cheaper, faster images

Fal released its own distilled take on the Flux 2 image generator, claiming roughly ten times lower cost and several times better efficiency than the base model. The ultra-fast variant is meant to beat much larger rivals on public image benchmarks.

Alibaba climbs to No. 2 in video models as Sora steps back

Alibaba's AI video model rose to second place in global rankings as OpenAI's Sora and ByteDance's Seedance lost ground. VentureBeat reports the field reshuffled sharply, with one major contender discontinued and another frozen amid rights disputes.

OpenCV's founders launch a video AI startup to take on the giants

The people behind OpenCV started a new venture focused on AI video generation, setting their sights on the offerings from OpenAI and Google. The move brings a longtime open computer-vision pedigree into the increasingly crowded text-to-video race.

Thinking Machines previews near-realtime voice and video chat

Thinking Machines showed an early look at new interaction models built for near-realtime conversation across voice and video. The demo points toward more fluid back-and-forth exchanges rather than the usual turn-based, lagging assistant experience.

Stability's enterprise audio model cuts 50 steps down to eight

Stability AI introduced an audio model aimed squarely at enterprise use that compresses generation from around fifty computational steps to just eight. The company says the shortcut slashes production time from weeks to minutes while keeping output quality high.

xAI's Grok 4.3 arrives cheap with a fast voice-cloning suite

xAI launched Grok 4.3 at an aggressively low price point and paired it with a new, quick voice-cloning toolkit. The bundle pushes the company further into audio alongside its broader model lineup.

Roblox opens up Cube, its generative 3D model for game makers

Roblox unveiled a set of generative AI tools at GDC, including an open release of its Cube 3D foundation model for building objects and scenes. Developers can run it on or off the Roblox platform and fine-tune it on their own data.

Yoroll.ai bets on world models for an engine-less game platform

Yoroll.ai is building what it calls the first engine-less game platform, leaning on world models that generate interactive worlds on the fly. VentureBeat frames it as part of a broader shift sparked by real-time interactive world models in gaming.

Salesforce launches Agentforce Operations to mend broken AI workflows

Salesforce introduced Agentforce Operations, a platform that breaks back-office processes into tasks handed to specialized agents via uploads or prebuilt blueprints. The pitch is that the real obstacle to enterprise AI is now the surrounding workflow plumbing, not the model.

Mistral's Workflows engine is already handling millions of runs a day

Mistral AI launched Workflows, a Temporal-powered orchestration engine for stringing together agent and model steps reliably. The company says the system is already processing millions of executions daily, underscoring how orchestration has become a core piece of enterprise AI.

OpenAI hands GPT-5.5-Cyber to a vetted circle of security defenders

OpenAI rolled out a hardened variant of GPT-5.5 that refuses far fewer offensive-security prompts and will actively run exploits against test machines. Access is restricted to verified people defending critical infrastructure, a guardrail meant to keep the dual-use model out of the wrong hands.

Google DeepMind fits a multimodal Gemma 4 model onto modest laptops

DeepMind released Gemma 4 12B, an open model that reads text, images, and audio natively without separate encoders and runs on machines with as little as 16 GB of RAM. The pitch is local multimodal AI that does not require a data center to be useful.

MiniMax ships an open-weight M3 model with a million-token window

Chinese lab MiniMax put out M3, which it calls the first open-weight release to pair top-tier coding ability with a one-million-token context and native multimodality. The combination is aimed squarely at the proprietary frontier models that have dominated those areas.

OpenAI's models and Codex land on Amazon Bedrock

OpenAI made GPT-5.5, GPT-5.4, and its Codex coding tool available through Amazon Web Services via Bedrock, covering both commercial and government cloud regions. The move widens where teams can run the models without going through OpenAI's own endpoints.

xAI enters the coding-agent race with terminal tool Grok Build

xAI launched Grok Build, its first command-line coding agent, stepping into a market that Anthropic carved out with Claude Code and OpenAI grew with Codex. The release is a late but direct challenge to the established terminal agents.

OpenAI turns Codex into a screen-watching, always-on assistant

Codex picked up a background computer-use mode that lets it see the screen, click, and type on its own, plus the ability to schedule its own future tasks. OpenAI says it can keep grinding on long projects across days or weeks without constant prompting.

Deepseek spins up a Beijing team to build a rival code agent

Deepseek is forming a new Harness group to develop its own coding agent from scratch, working under the name Deepseek Code. The effort takes direct aim at Claude Code and OpenAI's Codex from the open-model side.

Devin maker Cognition more than doubles its valuation to $26 billion

Cognition, the company behind the Devin coding agent, raised over a billion dollars at a valuation north of $26 billion. The round arrived in under nine months and underscores how hot investor appetite for autonomous coding tools remains.

ChatGPT Images 2.0 reasons and searches the web before drawing

OpenAI's updated image model spends time thinking ahead of generation and can even pull in web search during the process. It handles text in pictures, especially non-Latin scripts, far better and can produce up to eight consistent images from one prompt.

Alibaba's Qwen-Image-2.0 cuts generation from forty steps to four

Alibaba released Qwen-Image-2.0 with double the compression and a distilled variant that needs just four denoising steps instead of forty. The result is much faster image generation without the usual heavy step count.

Microsoft's Lens shows rich captions beat raw scale for image models

Microsoft Research unveiled Lens, a text-to-image model that competes with much larger rivals while using a fraction of the training compute. It needs roughly one-fifth the pre-training compute of comparable systems, with detailed captions doing the heavy lifting.

Microsoft's Mirage gives generated video a lasting sense of place

Mirage, a video world model from Microsoft Research and several universities, keeps scene geometry consistent even through long camera moves so the model does not forget what was just off-screen. It also runs up to 10.5 times faster and uses up to 55 times less memory than comparable systems.

xAI's Grok Imagine 1.5 turns stills into 720p clips

xAI updated Grok Imagine to version 1.5 with an image-to-video preview that animates a single still into a short clip at up to 720p. It is the company's push to keep pace in the fast-moving video generation field.

New benchmark says AI video looks great but still cannot reason about reality

A Tsinghua University benchmark tested whether top generators like Sora 2, Seedance 2.0, and Veo 3.1 can continue a scene in ways that make physical, social, and logical sense. The finding: stunning visuals and genuine world understanding remain two separate things.

Stability AI releases Stable Audio 3.0 with six-minute open-weight tracks

Stable Audio 3.0 can generate songs up to six minutes long and was trained entirely on licensed data. Three of the four model variants ship as open weights, with only the largest reserved for API and enterprise customers.

ElevenLabs Music v2 jumps from opera to metal without falling apart

ElevenLabs shipped Music v2 with sharper vocals, instrumentation, and arrangements across genres. A single track can swing between opera and heavy metal, handle fast rap, and weave in sound effects while staying musically coherent.

Odyssey's Agora-1 reimagines GoldenEye as a four-player AI simulation

AI lab Odyssey released Agora-1, a world model that lets up to four people share an AI-generated take on the N64 classic GoldenEye at the same time. It splits the work in two: one model tracks the shared game state while another renders each player's view live.

Amazon, Nvidia, and AMD pour $310 million into a 3D world-model startup

The venture arms of Amazon, Nvidia, and AMD together backed Odyssey ML with $310 million to build models that simulate the physical world in 3D. The bet signals heavy chip-and-cloud interest in interactive, generated environments.

OpenAI turns ChatGPT into a team automation platform with workspace agents

OpenAI began rolling out Codex-powered workspace agents that handle multi-step team workflows and keep running even when nobody is watching. The feature is in research preview for Business, Enterprise, Edu, and Teachers plans.

Anthropic launches Claude Fable 5 and Mythos 5 with big coding and science gains

Anthropic introduced two fifth-generation Claude models, with Fable 5 topping nearly every benchmark and Mythos 5 limited to select partners. On the SWE-Bench Pro test for real GitHub engineering tasks, Fable 5 reached 80.3 percent, ahead of Opus 4.8, GPT-5.5, and Gemini 3.1 Pro.

Ex-Datadog engineers start Niteshift to keep code out of model makers' hands

Two early Datadog engineers raised a $7M seed for Niteshift, an AI coding startup built on the idea that teams should not hand their source straight to the same labs that ship rival products. Greylock led the round, with angels including Reid Hoffman and Datadog cofounder Olivier Pomel.

Microsoft proposes an open standard for reining in AI agents

Microsoft introduced an open source effort called the Agent Control Specification, meant to give developers a uniform, fine-grained way to define what an autonomous agent is and is not permitted to do. The goal is more predictable behavior as agents take on real tasks across tools.

NewCore exits stealth with $66M to give AI agents real identities

Security startup NewCore came out of stealth with $66M to tackle how companies authenticate and govern AI agents that increasingly act like staff. Its package plugs into coding assistants such as Claude Code, OpenAI's Codex, and Cursor, with the seed round led by Cyberstarts.

Cognition chief argues coding agents should assist, not replace, engineers

Cognition CEO Scott Wu pushed back on the idea that AI coding agents will swap out human developers, framing them instead as a force multiplier for the people writing software. His comments land as agent tools rapidly fold into everyday engineering work.

Uber puts a monthly lid on employee spending for AI coding tools

After burning through a full year's AI budget in roughly four months, Uber set a $1,500 per-employee monthly cap on agentic coding tools like Claude Code and Cursor. The move is an early sign that runaway usage costs are forcing big companies to add guardrails.

Probably lands $9M to bolt a validator onto unreliable AI answers

Backed by Andreessen Horowitz, Probably raised $9M to make AI outputs far more dependable, starting with a data science tool. The system runs a model's first-pass answer through a deterministic checker, chasing the near-perfect accuracy people expect from traditional software.

New image models, not chatbot upgrades, are fueling AI app downloads

An Appfigures report found that image model launches drive far more app installs than routine chatbot updates, by a factor of roughly 6.5x. Both ChatGPT and Gemini picked up tens of millions of fresh downloads after shipping their image features.

Google jumps into AI design with a new Workspace app called Pics

At I/O 2026, Google unveiled Pics, an AI design and image app for Workspace that turns plain text prompts into graphics, invitations, marketing assets, and mockups. It debuted to testers at the event and is set to reach AI Ultra subscribers over the summer.

Avataar's Varya targets India with cheap, locally aware AI video

Indian startup Avataar launched Varya, a video model tuned to recognize regional festivals, food, and clothing while undercutting rivals on price. It generates a five-second 720p clip in about 45 seconds and will ship as an open-weight model on India's AIKosh portal.

Stability AI's new audio model can compose songs past six minutes

Stability AI rolled out a fresh family of audio models, with its largest able to produce coherent music running more than six minutes long. The medium and large versions hold structure and melody across full compositions, a step up from short generated clips.

Suno pulls in another $400M even as copyright suits hang over it

AI music maker Suno closed a $400M Series D that values the company at $5.4 billion, despite ongoing legal fights over how its system was trained. The raise follows Suno crossing 2 million paying subscribers earlier in the year.

ElevenLabs' Music v2 can flip genres in the middle of a track

ElevenLabs released Music v2, a generation model that can shift styles mid-song, jumping from opera to heavy metal and back without breaking. The company stressed the model was trained on licensed data and cleared for commercial use, a contrast to rivals tangled in lawsuits.

Roblox gives its AI assistant agentic powers to plan and test games

Roblox expanded its AI assistant with tools that can plan, build, and test games, including mesh generation that drops fully textured 3D objects into a world. The assistant grasps spatial relationships, letting creators place and scale objects with simple prompts.

Nvidia's DLSS 5 leans on generative AI to sharpen game visuals

Nvidia debuted DLSS 5, which blends conventional 3D rendering with generative models that predict and fill in parts of a frame to boost realism while cutting compute. The company signaled ambitions for the tech that reach beyond gaming.

Anthropic kicks off the AI IPO race with a public offering filing

Anthropic filed to go public, becoming one of the first frontier AI labs to start the journey to the stock market. The move came after a Series H that lifted its valuation toward the trillion-dollar range.

OpenAI files confidentially for an IPO, trailing Anthropic by a week

OpenAI submitted a confidential filing for an initial public offering just over a week after Anthropic did the same, sharpening the rivalry between the two labs. OpenAI carried a post-money valuation north of $850 billion from its prior round.

OpenAI recruits Transformer co-author Noam Shazeer before its IPO

Ahead of its public listing, OpenAI brought on Google DeepMind veteran Noam Shazeer, a Gemini co-lead and Character AI founder who co-wrote the 2017 paper that introduced the Transformer. It also hired former White House AI policy official Dean Ball.

Bezos-backed Prometheus raises $12B to build an AI engineer for hardware

Prometheus, co-founded by Jeff Bezos and Vik Bajaj, raised $12B at a $41 billion valuation to create what it calls an artificial general engineer for the physical world. The software aims to automate the design and manufacture of complex systems, from jet engines to drug compounds.

Jedify raises $24M to feed AI agents real business context

Jedify closed a $24M Series A led by Norwest to build a context graph that links into a company's knowledge sources through APIs. The idea is to give AI agents the grounding they need to handle real work without constant supervision.

Coralogix raises $200M to build a monitoring layer for AI agents

Boston-based Coralogix raised $200M in a Series F, wagering that the spread of autonomous agents will create demand for tools to watch, debug, and manage them. The bet is that someone has to keep tabs on increasingly self-directed software.

Reflection AI signs a SpaceX compute deal for Nvidia GB300 chips

Open source lab Reflection AI agreed to pay roughly $150M a month starting in July for access to Nvidia's latest GB300 chips and supporting hardware inside SpaceX's Colossus 2 data center near Memphis. The multi-year deal runs through 2029.

Google undercuts rivals by slashing its AI Plus subscription price

Google cut the monthly price of Google AI Plus from $7.99 to $4.99 while doubling included storage to 400GB, opening a front in the AI subscription price wars. The move pressures rivals charging more for comparable consumer plans.

Sandstone raises $30M to bring AI to in-house legal teams

Sandstone landed a $30M Series A led by Lightspeed to put AI tooling in the hands of corporate legal departments. The funding reflects steady investor appetite for vertical AI aimed at specific enterprise workflows.

VCs name the standout startups from Y Combinator's latest Demo Day

TechCrunch rounded up the 11 companies that drew the most enthusiasm from investors at Y Combinator's Demo Day. The list offers an early read on where the next wave of AI and developer-focused startups is heading.

Stealth startup says it cracked a math bottleneck slowing down LLMs

A company called Subquadratic emerged from stealth claiming it cut the number of computations transformers need to run, which it says yields a faster and cheaper model. If the work holds up, it could meaningfully lower the energy bill behind today's largest systems.

Three takeaways from DeepSeek's new flagship model

DeepSeek shipped a preview of V4, keeping the model open source while pushing its performance up against closed rivals from Anthropic, OpenAI, and Google. The pricing is the headline, with token costs running far below what Western labs charge.

Anthropic's developer event made the case for an agent-driven coding future

At its Code with Claude gathering, Anthropic leaned hard into a vision where coding agents do much of the heavy lifting. The demos suggested where the craft is heading, even for developers who are not thrilled about handing over the keyboard.

AI coding has gone mainstream, but plenty of developers remain unsold

Tools like Copilot, Cursor, and Replit have put app and website building within reach of people who barely write code. Even so, a chunk of working engineers question whether the output is reliable enough to trust without close human review.

Generative coding lands on the 2026 breakthrough technologies list

MIT Technology Review named generative coding one of its breakthrough technologies for the year, citing how quickly natural-language prompts now translate into working software. The piece tracks the jump from autocomplete helpers to agents that assemble entire programs.

Coordinating swarms of agents is becoming its own discipline

Agent orchestration made the list of ideas reshaping AI right now, as developers wire multiple agents together to chase goals no single model could finish alone. The shift moves the hard problem from one model's smarts to how a team of them divides the work.

OpenAI is betting big on an AI that can run research on its own

OpenAI is pouring resources into what it calls an AI researcher, an agent system meant to tackle large problems without human steering. The company has floated an autonomous research intern as a near-term step toward a fuller multi-agent setup later this decade.

AI co-scientists are starting to act like real collaborators

Labs and companies are building agents that hunt down prior results, propose hypotheses, and sketch out experiments to test them. Some researchers think these systems could eventually contribute work worthy of major scientific prizes.

World models move to the center of the AI conversation

Google DeepMind, World Labs, and others are building systems that spin up interactive 3D environments from text, images, and video. Beyond games and VR, the bigger prize is giving agents an internal map of the world so they can predict the results of their actions.

Yann LeCun's new lab is a bet against the LLM playbook

After leaving Meta, Yann LeCun launched a venture built on the idea that large language models are a dead end for real intelligence. His focus on world models puts him on a collision course with most of the industry's current direction.

Song-generating AI is muddying what authorship even means

New diffusion models can compose full tracks from a prompt, raising thorny questions about who, if anyone, counts as the artist. The technology is forcing a rethink of creativity and credit just as the music industry braces for the fallout.

AI art grows up as some creators reach galleries and auction houses

A handful of artists working with image generators are building large followings and selling pieces at auction. The story marks a turn from throwaway output toward work that the traditional art world is starting to take seriously.

The personal toll of finding yourself in nonconsensual deepfakes

MIT Technology Review reported on what it feels like to discover your likeness used in AI-generated explicit videos, and the grinding fight to get them removed. The piece digs into the gaps in takedown systems and copyright law that leave victims exposed.

Capturing human motion at scale is fueling the humanoid robot push

Companies are recording huge volumes of everyday human movement to teach humanoid robots tasks like wiping tables and stacking dishes. With billions flowing into the sector, gathering that training data has spawned a sprawling new gig economy.

The hidden human labor propping up humanoid robot demos

Behind many impressive robot demonstrations sits remote teleoperation and concealed human effort that companies rarely disclose. The opacity makes it easy to mistake staged routines for genuine machine autonomy.

Pokemon Go data is giving delivery robots a sharper view of streets

Map data gathered through the AR game is helping delivery robots navigate sidewalks with surprising precision. It is an unexpected example of consumer play feeding directly into the spatial models that robots rely on.

Courts are drowning in filings written with AI

Judges are grappling with a surge of legal documents drafted with chatbots, including fabricated citations that slip into the record. The trend is straining a court system that was never built to vet machine-generated arguments at this volume.

Flexible power deals could get AI data centers online far sooner

A trial run by a tech, utility, and grid-operator coalition showed server racks dialing back their draw when the grid tightens. Letting data centers flex demand could shave years off the wait to connect new sites to the power supply.

Virtual power plants pitched as a fix for hungry data centers

A US proposal would let new data centers connect to the grid years earlier if they agree to cut demand when supply runs short. The approach treats clusters of flexible load as a kind of distributed power resource rather than a pure drain.

Five things worth knowing about AI right now

MIT Technology Review distilled the current AI moment into a short briefing on where the field actually stands. It cuts through hype cycles to flag the developments most likely to matter in the months ahead.

Visual Studio 2026 leans harder into AI-assisted development

Microsoft's latest Visual Studio release pushes AI deeper into the editor, including a Profiler Agent that helps developers track down and fix performance problems without being profiling experts. The update positions AI assistance as a default part of the everyday coding workflow rather than an add-on.

OpenAI ships a standalone Codex app for managing coding agents

OpenAI released a dedicated Codex application that lets developers coordinate multiple AI coding agents across projects, moving past simple chat-driven code generation. The launch arrives as companies debate how much autonomy to grant these tools and how to govern them.

JetBrains introduces a control layer for coding agents called Central

JetBrains unveiled JetBrains Central, an agentic development platform that gives teams oversight and management across their AI coding agents. It works alongside JetBrains Air and the model-agnostic Junie agent, with an early access program slated to begin in the second quarter of 2026.

Google folds its AI coding tools into a single Antigravity platform

Google is consolidating its developer AI tools under Antigravity, aiming to support the full agentic software lifecycle instead of offering disconnected assistants. The idea is a persistent layer where project context, run history, and agent state carry across coding, testing, debugging, and deployment.

Six AI advances expected to shape the rest of 2026

InfoWorld lays out the developments it expects to define the year, including self-verifying agents that catch and correct their own mistakes mid-task. The piece frames error accumulation in multi-step workflows as the main barrier that needs solving before agents can scale in the enterprise.

Small language models force a rethink of enterprise AI design

The case for compact models is growing as teams weigh speed, cost, privacy, and lower resource demands against the heavyweight LLM approach. InfoWorld argues this shift is changing how architects plan AI systems rather than just trimming a few parameters.

A look at 21 language models built for specific industries

InfoWorld surveys specialized models tuned for narrow fields, from medicine to law to finance, including Microsoft's PubMed-trained BioGPT and JPMorgan Chase's contract analysis system. The takeaway is that domain-specific tuning is becoming a serious alternative to relying on one general model for everything.

The 33 LLM metrics worth keeping a close eye on

This guide breaks down the measurements that matter when evaluating large language models, covering accuracy, cost, latency, and safety dimensions. It is aimed at teams trying to compare models with more rigor than vibes and benchmarks alone.

Multi-token prediction reportedly triples LLM inference speed

A new technique predicts several tokens at once and claims a roughly threefold speedup in inference without needing separate draft models. If it holds up, it could cut serving costs for production systems that currently lean on speculative decoding.

Gemini 2.5 Flash Image steps up Google's picture generation

Google's Gemini 2.5 Flash Image model brings faster generation and editing through the Gemini API, AI Studio, and Vertex AI for enterprise use. It targets developers who want to wire image creation and manipulation directly into their applications.

Microsoft's Phi-4 multimodal model takes on speech, text, and video

Microsoft's Phi-4-multimodal is a 5.6 billion parameter model that uses a mixture-of-LoRAs approach to handle speech, vision, and language together. The company pitches it as an efficient, scalable option for developers building multimodal features.

Google brings local agentic AI to laptops with Gemma 4 12B

Google released tooling that lets developers run agentic workflows on their own machines using the 12-billion-parameter Gemma 4 12B from DeepMind. The push is about keeping agent execution local rather than routing everything through paid cloud APIs.

Google's DiffusionGemma drops the left-to-right generation habit

DiffusionGemma is a 26B mixture-of-experts model built on the Gemma 4 family and Google's Gemini Diffusion research, generating text without strict sequential decoding. The design aims to boost text output throughput by breaking away from token-by-token generation.

Gemma 4 holds up well on local hardware, big and small

Google's Gemma 4 family spans several sizes and quantizations so it can run on everything from servers down to commodity PCs. InfoWorld's review highlights its reasoning, tool use, and multimodal vision and audio support as standout features for local deployment.

Why world models are being called AI's next big leap

InfoWorld digs into world models that learn the rules of reality from data and can run interactive, game-engine-free simulations. It points to examples like Decart's playable AI environments and World Labs' Marble, which rebuilds 3D scenes from still images and lets users reshape them on the fly.

Ten MCP servers that plug AI agents into DevOps work

InfoWorld rounds up Model Context Protocol servers that give AI agents real abilities across Git, CI/CD, infrastructure as code, observability, and documentation. The piece reflects how quickly MCP has become a connective standard for agent-driven operations.

Databricks tackles AI ops headaches with Genie ZeroOps

At its Data and AI Summit, Databricks introduced Genie ZeroOps, an agentic capability that automates monitoring, investigation, and remediation across data and AI workloads. The goal is to keep pipelines and models running without constant manual babysitting.

Databricks pitches OpenSharing to cut the AI integration tax

Databricks announced OpenSharing, an open protocol for sharing models, agent skills, dashboards, and unstructured data across platforms without copying or moving the assets. It targets the friction enterprises hit when stitching AI tooling together across systems.

AWS adds release management muscle to its DevOps Agent

AWS updated its DevOps Agent with features that automatically check code changes against company standards, flag release risks, and generate tailored tests before code ships. The aim is to clear bottlenecks in the path from commit to production.

First look at Mojo 1.0, a Python and Rust mashup

Mojo reached its 1.0 milestone with a language that compiles to native machine code while borrowing Rust-style memory safety. InfoWorld's hands-on covers how it blends Python familiarity with performance ambitions aimed at AI and systems work.

SpaceX agrees to acquire AI coding startup Cursor in a $60 billion stock deal

SpaceX has struck a deal to purchase Cursor, the company behind the popular AI-assisted code editor, for roughly $60 billion in stock. The agreement closes out talks that first surfaced earlier in the year and folds one of the best known vibe-coding tools into Elon Musk's rocket firm.