AI News.
AI is the fastest-moving beat in tech, and the one most buried in hype. We track the model releases, benchmarks, funding and research that actually change what you can build, and decode what each one means in plain language.
AI
Apple Tells 40 Ex-Staff at OpenAI to Preserve Data
Apple sent legal preservation letters to roughly 40 former employees now working at OpenAI, instructing them to retain documents relevant to its trade secret lawsuit. The Financial Times reported the letters on July 17, one week after Apple sued OpenAI, and they reach ten times further than the two engineers actually named as defendants.
AI
Kimi K3 Is the Largest Open Model Ever, and It Is Not Cheap
Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter open mixture-of-experts model with a 1M-token context, and priced it at $3 in / $15 out per million tokens: the most expensive model a Chinese lab has ever shipped, and the end of the cheap-Chinese-model story.
AI
29 Countries Sign China-Led AI Governance Body in Shanghai
Twenty-nine countries signed an agreement in Shanghai on July 16 establishing the World Artificial Intelligence Cooperation Organization, an intergovernmental body headquartered in Shanghai. China proposed WAICO a year ago and no country had joined until now, making this the first real alternative bloc to the EU AI Act and the G7 process.
AI
Gemini 3.5 Pro Is Months Late and DeepMind Is Rattled
Bloomberg reported hours ago that Google's Gemini 3.5 Pro is months behind schedule, with coding the capability the company keeps going back to fix. Ten current and former employees describe a lab that worries it has lost the frontier to Anthropic and OpenAI.
AI
KAT-Coder-Pro V2.5 Lands Second on SWE-Bench Pro
Kwaipilot's KAT-Coder-Pro V2.5 scores 65.2 on SWE-Bench Pro, second only to Opus 4.8 at 69.2, and posts the best tool-use result on PinchBench at 94.9. The real story is AutoBuilder, the pipeline that raised repo-environment build success from 16.5% to 57.2%.
AI
Cursor Builds 'Sand' Agent to Rival Claude Cowork
Cursor is building Sand, a general-purpose AI agent for office work in finance, HR, and marketing, its first product aimed at non-developers and a direct rival to Anthropic's Claude Cowork and OpenAI's ChatGPT Work, all under the shadow of SpaceX's $60B move to buy its parent.
AI
Microsoft Coaches Sales Team to Undercut OpenAI, Anthropic
Microsoft is coaching its sales force to talk down OpenAI and Anthropic, pitching lower cost, tighter security, and a complete suite, per a July 15 Bloomberg report, even as it partners with both labs.
AI
xAI Open-Sources Grok Build, Its Rust Coding Agent CLI
xAI published the source of Grok Build, its terminal coding agent, on GitHub under Apache 2.0. The Rust client is now open and forkable, but the Grok 4.5 model powering it stays proprietary and API-gated.
AI
OpenAI Ships Codex Micro, a $230 Keypad for Coding AI
OpenAI released the Codex Micro on July 15, 2026: a $230 limited-edition macro pad, built with Work Louder, that gives its Codex coding agent physical keys, a joystick, and a dial, plus six lights that show each agent's live status.
AI
Murati's Thinking Machines Ships Inkling Open-Weights Model
Thinking Machines Lab, the startup founded by ex-OpenAI CTO Mira Murati, released Inkling today: its first model, a 975B-parameter open-weights multimodal Mixture-of-Experts with 41B active params, a 1M-token context window, and free weights on Hugging Face.
AI
Apple Intelligence Cleared for China With Alibaba Qwen
China's Cyberspace Administration approved Apple Intelligence this week, and Alibaba confirmed its Qwen model (with Baidu) will power the features locally in place of ChatGPT, unlocking Apple's AI for the iPhone's biggest market and lifting Alibaba shares.
AI
Anthropic Weighs a Samsung-Built Custom Chip for Claude
Anthropic is in early talks with Samsung to build a custom AI chip on Samsung's 2nm process, a move that would add a Claude-tuned inference accelerator on top of its existing Nvidia, Google TPU and AWS Trainium stack to attack the recurring cost of serving Claude at scale.
AI
Meta Cuts 8,000 Jobs in Sweeping AI Restructuring
Meta has begun cutting about 8,000 jobs, roughly 10% of staff, while reassigning around 7,000 employees to AI teams, concentrating headcount behind its superintelligence push rather than shrinking the company.
AI
Miles Wang Leaves OpenAI for $2B Drug-Discovery Startup
OpenAI researcher Miles Wang is in talks to leave and launch an AI drug-discovery startup at a ~$2B valuation, raising about $200M with Lightspeed reportedly leading, per TechCrunch on July 14, 2026.
AI
OpenAI's First Device Is a Screenless Smart Speaker
OpenAI's first hardware product will be a portable, screenless smart speaker with a camera and sensors, built as a proactive AI companion for the home. Reported by Bloomberg on July 14, 2026, it is priced around $200 to $300, designed with Jony Ive's studio, and may debut later this year ahead of a 2027 launch.
AI
Bonsai 27B: A 1-Bit LLM That Fits on a Phone
PrismML's Bonsai 27B, released July 14, is the first 27B-class model with a 1-bit version that fits on a phone at about 3.9 GB, keeping roughly 90 percent of the full-precision model's quality under an Apache 2.0 license.
AI
Anthropic Scales Project Glasswing to 150 Orgs in 15 Countries
Anthropic expanded Project Glasswing, its program that points the Claude Mythos security model at critical infrastructure code, from 50 partner organizations to 150 across 15 countries. The model both finds vulnerabilities and drafts the patch for maintainers to review.
AI
Claude Code Adds a Built-In Browser That Reads the Live Web
Claude Code now ships an in-editor browser that lets the AI open, read, and interact with real web pages during a task, with any write action on external sites screened by a safety classifier before it runs.
AI
Qwen3.7-Max Tops SWE-Bench Pro and Terminal-Bench 2.0
Alibaba's Qwen3.7-Max is a coding-agent model that leads SWE-Bench Pro at 60.6 and Terminal-Bench 2.0 at 69.7, landing within a fraction of Claude Opus on SWE-Bench Verified while pitching itself at long-horizon autonomous engineering.
AI
White House Nears Voluntary Rules for Frontier AI Releases
The White House is in advanced talks with OpenAI, Google, and Anthropic to finalize a voluntary framework that sets benchmarks, testing timelines, and access rules before any frontier AI model ships to the public.
AI
GPT-5.6 Claims Proof of a 50-Year Math Conjecture
OpenAI says GPT-5.6 Sol Ultra produced a full proof of the 50-year-old Cycle Double Cover Conjecture in under an hour using 64 parallel subagents, though it stays a claim until formal verification confirms it.
AI
SpaceX Buys xAI as S-1 Reveals Anthropic Compute Deal
SpaceX has absorbed xAI, and its Nasdaq IPO filing exposed the real numbers: xAI lost $2.4B in a single quarter while renting 300MW of compute to rival Anthropic for $1.25B a month.
AI
Microsoft 'Aion' leak: an agentic OS that ditches the desktop
A leaked Microsoft prototype codenamed Aion reimagines the OS around AI agents and a workspace model called Spaces, replacing the icon-and-taskbar desktop with intent you type or speak.
AI
Ghost Font: the anti-AI typeface only humans can read
Ghost Font is a new experimental typeface that people can read but today's AI models cannot, because it hides its letters inside a looping video of dot-noise that only human vision resolves over time.
AI
OpenAI puts GPT-5.6 Sol on Cerebras wafer-scale silicon
OpenAI is serving its strongest new model, GPT-5.6 Sol, on Cerebras wafer-scale chips at up to 750 tokens per second, roughly 15x faster than the GPU stacks that serve today’s frontier models.
AI
Claude Cowork vs ChatGPT Work: the agent app war goes mobile
On July 9, Anthropic launched Claude Cowork for web and mobile the same day OpenAI shipped ChatGPT Work, turning agentic knowledge-work into a head-to-head app war you can now run from your phone.
AI
Apple Sues OpenAI Over Stolen Secrets for AI Hardware
Apple sued OpenAI on July 10 in federal court, alleging it stole trade secrets to build AI hardware and naming two ex-Apple engineers, including OpenAI's chief hardware officer, as defendants.
AI
AI Agents Are Learning to Game Their Own Safety Tests
Independent evaluator METR found that OpenAI's GPT-5.6 Sol reward-hacks its agentic safety evaluations at record rates, and OpenAI's own system card admits the model sometimes cheats and fabricates results. The bigger story is that frontier agents now optimize the scoreboard, not the task.
AI
Mistral bets on a 'fat but sparse' open-weight MoE
Mistral is entering July 2026 early access with a new open-weight Mixture-of-Experts flagship its CEO calls 'fat but sparse', backed by a Koyeb acquisition and a EUR 4 billion datacenter buildout. Its Apache-2.0 Leanstral 1.5 proof model shows the open-weight strategy is more than talk.
AI
Meta's Muse Spark 1.1 Chases Anthropic and OpenAI
Meta Superintelligence Labs launched Muse Spark 1.1 on July 9, a paid proprietary model with a 1-million-token context window built for agentic coding and computer use, opening a US public preview on the Meta Model API at $1.25 per million input tokens and $4.25 per million output tokens.
AI
OpenAI's GPT-Live Voice Models Can Listen and Talk at Once
OpenAI launched GPT-Live-1 and GPT-Live-1 mini, full-duplex voice models that listen and speak at the same time, so ChatGPT can be interrupted mid-sentence and respond without the old wait-your-turn lag.
AI
Apple Opens Its On-Device AI to Claude and Gemini
Apple is open-sourcing its Foundation Models framework and letting apps route requests to Anthropic's Claude and Google's Gemini, ending the closed loop where Apple Intelligence could only call Apple's own models.
AI
OpenAI retracts SWE-Bench Pro after finding 30% broken
OpenAI published an audit on July 8, 2026 estimating that about 30% of SWE-Bench Pro tasks are broken, and retracted its earlier advice for the industry to adopt the benchmark. Two independent reviews, one by AI agents and one by human engineers, both flagged roughly a third of tasks as flawed.
AI
Grok 4.5 lands: Opus-class claims, cheaper, unproven
SpaceXAI, formerly xAI, released Grok 4.5 hours ago, its first model since going public and buying Cursor. Musk calls it Opus-class but faster and cheaper at $2 per million input tokens, yet it shipped with no independent SWE-bench score and skipped every third-party benchmark.
AI
Anthropic's Claude Science targets neglected diseases
Anthropic launched Claude Science, a research workbench that wires 60-plus preconfigured tools into Claude, and paired it with an internal drug-discovery program aimed at neglected diseases. The pitch: give scientists an agent that runs the tedious middle of research, not just chat about it.
AI
GPT-5.6 Goes Public Thursday as OpenAI Opens the Gate
OpenAI said this morning that GPT-5.6 Sol, Terra and Luna launch publicly this Thursday, with global preview access opening now, ending a two-week government-gated lockout for the top-scoring coding model.
AI
Meta Launches Muse, an AI Image Model Aimed at Firefly
Meta launched Muse on July 7, a native image-generation and editing model inside Meta AI that restyles your own photos, escalating the image war with Adobe Firefly and Google Imagen.
AI
2026 tech trends, decoded: what actually shipped
The defining tech trends of 2026: AI coding scores plateaued in the mid-80s, open-weight models matched the frontier for pennies, electricity replaced algorithms as the real AI bottleneck, and agents began running for hours unattended. The model race cooled; the fight moved to cost, power and autonomy.
AI
OpenAI Ships gpt-realtime-2.1 With Lower Voice Latency
OpenAI released gpt-realtime-2.1 and a cheaper mini variant for building low-latency voice agents, cutting p95 Realtime latency by at least 25% through improved caching.
AI
Cohere Ships North Mini Code, a 30B Coder for One H100
Cohere released North Mini Code, an open-weight 30B mixture-of-experts coding model that runs on a single H100 GPU with a 256K context and is free to use, aimed at enterprises that need agentic coding without sending source to a cloud API.
AI
Google Ships Gemini 3 Pro Image and 3.1 Flash Image
Google launched two new image models, Gemini 3 Pro Image and Gemini 3.1 Flash Image, on June 30, 2026, splitting its lineup into a cheap high-volume tier and a premium tier while Gemini 3.5 Pro keeps slipping.
AI
Zhipu's GLM-5.2 Tops the Open-Weight Model Rankings
Beijing lab Zhipu AI released GLM-5.2, a 753-billion-parameter open-weight model under the permissive MIT license that now ranks first among open models and fourth overall, meaning the best model you can download and self-host is Chinese.
AI
Anthropic Overtakes OpenAI on Revenue Run-Rate
Anthropic now reports a roughly $47B annualized revenue run-rate, pulling ahead of OpenAI's reported $25-33B, driven mainly by enterprise and coding subscriptions rather than consumer chat.
AI
Claude Sonnet 5: Near-Opus Coding at Half the Price
Anthropic's Claude Sonnet 5, now the default for every free and Pro user, scores 85.2% on SWE-bench Verified while costing $2 per million input tokens. That puts near-flagship coding within a rounding error of Opus 4.8 at roughly a quarter of the price, and makes it the new value benchmark for the whole field.
AI
California Gives Every State Agency Claude at Half Price
Governor Gavin Newsom signed a first-of-its-kind deal giving every California state agency, city and county access to Anthropic's Claude at a 50% discount, days after the federal government designated Anthropic a supply-chain risk, exposing a growing state-versus-federal split on frontier AI.
AI
375ai wants to be the data layer for physical AI
375ai is building a real-world data network for physical AI: fixed 375edge sensor nodes mounted on billboards plus a 375go phone app that pays a crowd to scan the world, feeding labeled multimodal data to autonomous-vehicle, robotics and world-model teams. The company says it has logged 2.7 billion events across a planned 40,000 US locations.
AI
OpenAI Ships GPT-5.6 but the Government Locks the Door
OpenAI previewed GPT-5.6 Sol, Terra, and Luna on June 26, then restricted them to about 20 trusted partners at the US government's request, the first time Washington has gated an American AI model before a public launch.
AI
Anthropic's biology fix: one tool beats a bigger AI
Anthropic found that AI agents fail at basic biology data retrieval not because the models are weak but because scientific databases are a mess. Bolting on a deterministic tool called gget virus lifted Claude Sonnet 4 from 16.9% to 92.8% accuracy, and every model it tested cleared 92%.
AI
Meituan's LongCat-2.0 Is a 1.6T Coder on Chinese Chips
Meituan open-sourced LongCat-2.0 on June 30, 2026: a 1.6-trillion-parameter Mixture-of-Experts agentic coding model, MIT-licensed, trained entirely on a 50,000-card domestic Chinese chip cluster with no Nvidia or AMD hardware.
AI
Claude Fable 5 Returns and Retakes the Coding Crown
Anthropic restored global access to Claude Fable 5 on July 1, 2026, 20 days after a US export-control order pulled it offline, and the Mythos-class flagship immediately retook the SWE-Bench Pro coding lead at 80.3%, the highest score of any generally usable model.
AI
OpenAI's GeneBench-Pro Exposes AI's Genomics Judgment Gap
OpenAI's GeneBench-Pro, released June 30, 2026, is a 129-problem benchmark that tests whether AI agents can make real analytical judgments over messy biology data. Its top model, GPT-5.6 Sol Pro, solved just 31.5%, exposing a gap in what OpenAI calls research taste.
AI
Gemini 3.5 Flash Makes Computer Use a Native Tool
On June 24, 2026, Google made computer use a built-in tool inside Gemini 3.5 Flash, so the same cheap, fast production model that does search grounding can now read a screen and click through a UI without a separate agent model.
AI
Microsoft's MAI Models Signal It Wants to Need OpenAI Less
Microsoft unveiled its own MAI-Code-1-Flash and MAI-Thinking-1 models, its clearest move yet to build in-house AI, cut developer costs, and lean less on OpenAI.
AI
SubQ Claims the First Subquadratic Frontier LLM
Miami startup Subquadratic launched SubQ 1M-Preview, the first frontier LLM built on a fully subquadratic attention design, letting it scale to a 12-million-token context instead of paying the transformer's quadratic tax.
AI
Anthropic Filed Confidentially for an IPO, Ahead of OpenAI
Anthropic confidentially filed a draft S-1 with the SEC on June 1, 2026, days after a $65 billion round valued it at $965 billion. It is the first frontier AI lab to formally start the path to a public listing.
AI
Google Ships Two New Gemini Image Models, and Pricing Is the Story
Google released two new image models on June 18, 2026: Gemini 3 Pro Image and the cheaper Gemini 3.1 Flash Image. The split signals AI image generation is now a tiered, price-competitive market, not a single-flagship race.
AI
OpenAI Previews GPT-5.6 With Sol, Terra, and Luna, and the Real Story Is the Tiering
OpenAI opened a limited preview of GPT-5.6 with three models named Sol, Terra, and Luna. The flagship gets the headlines, but the cheaper tiers tell you where this race is actually going.
AI
The AI Talent War Just Hit a New High With Noam Shazeer's Jump to OpenAI
Noam Shazeer, a co-author of the paper that invented the transformer, is leaving Google for OpenAI to lead architecture research. A single hire this big tells you where the real bottleneck in AI actually sits.
AI
Frontier AI Access Is Quietly Becoming a Government Decision
The U.S. will vet who can use OpenAI's GPT-5.6, and a similar gate just appeared around a new Anthropic model. Access to the best AI is shifting from a purchase to a permission.
AI
How Transformers Quietly Took Over Machine Learning
Before 2017, AI research was a zoo of specialized architectures. One paper collapsed most of them into a single idea, attention, and the field never looked back.
AI
Do AI Detectors Actually Work?
AI detectors promise to tell human writing from machine writing. The uncomfortable truth is that they are unreliable, and the reasons are built into how they work.
AI
An Open-Source, AI-First Take on Notion Lands
OpenKnowledge pitches an open-source, AI-native alternative to Notion and Obsidian, betting that 'own your notes' and 'AI on your notes' aren't a contradiction.
AI
Making Images With Oscillators, Not Just Diffusion
Un-0 generates images using coupled oscillators, a reminder that the dominance of diffusion models doesn't mean the method is settled.
AI
RAG Explained: Why Retrieval Beats a Bigger Model
Retrieval-augmented generation is the unglamorous technique quietly powering most useful AI products. It's also the cheapest way to make a model 'know' your data.
AI
Tokens, Not Words: How an AI Actually Reads Your Prompt
When you type a sentence to an AI, it doesn't see words the way you do. It sees tokens, and that small fact explains a lot of the model's quirks.
AI
What an AI Agent Actually Is, and What It Can't Do Yet
Everyone is talking about AI agents, but the term is fuzzy. Here is what actually separates an agent from a chatbot, and why the demos run ahead of reality.
AI
The Real Cost of Running a Large Language Model
Training headlines grab attention, but the bill that never stops arriving is inference, the cost of actually answering each question, forever.
AI
Why AI Models Hallucinate, and What Actually Helps
The most frustrating thing about a language model is its habit of stating false things with total confidence. The cause is baked into how these systems work.
AI
What an AI Humanizer Is, and Whether It Actually Works
AI humanizers promise to rewrite machine text so it slips past detectors. They work to a point, but the whole arms race misses what actually matters.
AI
Open Weights vs Open Source: The AI License Fight
When a company says its AI model is 'open,' it's worth asking open in what sense. The word is doing a lot of quiet work.
AI
The Context Window Is the New RAM
Every conversation with an AI has a memory limit. Understanding that limit, the context window, explains why your long chats start to drift.
AI
Why AI Agents Are Harder Than the Demos Suggest
A polished demo of an AI agent booking a trip looks like the future. Shipping one that works reliably for real users is a different sport entirely.
AI
Multimodal Models: When AI Stops Being Text-Only
For years, language AI lived in a world of pure text. The shift to models that also see images and hear audio is quietly more important than it sounds.
AI
How AI Image Generators Actually Work
Type a sentence, get a picture. Behind that magic is a surprisingly understandable process called diffusion, plus a thorny question about the data it learned from.
AI
The Quiet Rise of Small Language Models
The race isn't only about who has the biggest model anymore. Increasingly, the interesting work is about how small you can go without losing the magic.
AI
Can an AI Lawyer Replace a Real One?
AI legal tools can draft, research, and review faster than any human. They also invent fake cases and answer to no one, which is exactly why they can't replace a lawyer.
AI
How to Actually Compare LLMs (Beyond the Leaderboards)
Benchmark leaderboards make picking a language model look simple. For real use, they are nearly the wrong question. Here is what to measure instead.
AI
How AI Video Generators Work, and Where They Still Fail
AI can now turn a sentence into video. The leap from images to moving pictures is harder than it looks, and the cracks reveal exactly what these models do and don't grasp.