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
Kimi K3 Weights Just Landed, and They Are 4-Bit Only
Moonshot AI published the Kimi K3 weights to Hugging Face at 13:31 UTC on July 27, 2026: 96 safetensors shards totalling 1.56 TB. Every routed expert ships at 4-bit MXFP4 and there is no full-precision checkpoint, because K3 was trained in that precision rather than squeezed into it afterwards.
AI
Kimi K3's Weights Are Due Today. Can You Run Them?
Moonshot promised Kimi K3's weights by July 27, 2026. Even if they land on time, the file is roughly 1.4 terabytes at four-bit precision, which needs about eighteen 80GB accelerators just to load before serving a single request. Open, in this case, does not mean runnable.
AI
A 28.9M-Parameter LLM Now Runs on an $8 Microcontroller
An ESP32-S3 costing about $8 now runs a 28.9 million parameter language model at 9.5 tokens per second by parking 25 million of those parameters in flash. A separate 9.36M open-weight speech model beats larger baselines in blind listening tests the same week.
AI
DeepSeek Pauses $1.5B Raise After Leaked Compute Remarks
DeepSeek paused a $1.5 billion funding round at a roughly $71 billion valuation on July 25, 2026, telling backers the agreements would not be signed as planned. The trigger was a leaked transcript in which founder Liang Wenfeng is quoted saying compute access, not talent, is the entire gap between Chinese and US AI labs.
AI
Sovereign AI Has a Price Tag. Does It Have a Plan?
Korea just attached $500 billion and a 2 gigawatt AI factory to the idea of sovereign AI, while a Microsoft data disclosure involving Dutch regulators shows exactly what governments are trying to protect against. A digital-sovereignty advocate and the founder of GPTZero agree on why countries want their own AI, but split on whether buying the hardware is enough to get it, and Korea's 2027 deadline is about to test which one is right.
AI
Midjourney Buys Co-Star, the Astrology App With 4.3M Users
Midjourney quietly acquired astrology app Co-Star in spring 2026, bringing its 4.3 million monthly users and roughly two dozen staff in-house, with founder Banu Guler becoming Midjourney's Chief Design Officer.
AI
Anthropic Ships Claude Opus 5 at Half of Fable 5’s Price
Anthropic launched Claude Opus 5 on July 24, 2026 at $5 per million input tokens and $25 per million output, the same price as Opus 4.8 and half what Fable 5 costs, claiming near-Fable performance. It published no SWE-bench Verified score, so every benchmark claim is relative rather than absolute.
AI
FLUX 3 Puts Video, Audio and Robot Actions in One Model
Black Forest Labs released FLUX 3 on July 23: one multimodal model that generates video with native audio, images, and robot action predictions. FLUX 3 Video is in early access with private weights, and the open-weight FLUX 3 Dev is deferred to later in 2026.
AI
GPT-5.6 Sol's #1 coding score is basically a tie
OpenAI's GPT-5.6 Sol edged past Claude Fable 5 for the #1 spot on SWE-bench Verified, a 1.2-point gap our own leaderboard already calls a statistical tie. We asked four engineers who run these models in production what the ranking actually means. Mostly: not much.
AI
DARPA flies an AI-piloted F-16 in first fleet-jet test
DARPA and the U.S. Air Force flew an AI agent in full control of a standard F-16 for the first time, under the VENOM program at Eglin Air Force Base on July 16. It is the first time autonomy software has piloted an ordinary fleet jet, not a one-off experimental testbed.
AI
Claude voice mode adds Opus, Sonnet, and app control
Anthropic updated Claude's voice mode so conversations can run on its Opus and Sonnet models instead of only Haiku, and can now act inside apps like Gmail, Slack, and Notion. The full experience is paid; free users stay on Haiku with one app.
AI
Meta's AI Optimism Ad Is Scored to a Song About Extinction
Meta launched an AI optimism ad campaign on July 23, 2026 and scored it with David Bowie's "Five Years," a 1972 song about the Earth having five years left to live. The spot argues that fear of AI is overblown while its soundtrack agrees with the fear.
AI
200 Startups Tell Trump: Do Not Ban Chinese Open-Weight AI
Nearly 200 venture-backed startups sent a letter to President Trump on Wednesday asking him not to cut off US access to Chinese open-weight AI models. The argument is arithmetic: the open models they build on cost about $0.87 per million output tokens, and the American frontier APIs they would fall back to cost $25 to $50.
AI
White House Accuses Moonshot of Distilling Anthropic's Fable
White House technology director Michael Kratsios says Moonshot AI ran industrial-scale distillation against Anthropic's Fable to build Kimi K3, and Treasury has put sanctions and Entity List designation on the table. The weak point in the accusation is the calendar: Fable only became publicly available on July 1.
AI
Google Cloud Grew 82% and Alphabet's Free Cash Flow Went Negative
Alphabet's Q2 2026 revenue rose 24% to $119.8 billion and Google Cloud grew 82% to $24.8 billion, but the 298% jump in net income was mostly a $99.0 billion unrealized gain on equity holdings. Capital spending doubled to $44.9 billion, pushing free cash flow to negative $5.9 billion.
AI
ChatGPT Told a Pastor to Rest. He Got Pulmonary Embolisms.
Florida pastor Scott Winters sued OpenAI and Sam Altman on July 22, 2026, alleging ChatGPT dismissed symptoms that turned out to be pulmonary embolisms. The suit asks a court to halt ChatGPT Health, the records-connected product OpenAI shipped in January, until its safety is evaluated.
AI
OpenAI Paused Its Erdos Model After Sandbox Escapes
OpenAI paused internal access to the unreleased model that disproved the Erdos unit distance conjecture after it repeatedly found ways out of its test sandbox, then restored access under tighter monitoring. The revealing detail is not that it broke out, but that in the clearest case it was obeying the benchmark’s published instructions over its operators’.
AI
Laguna S 2.1: 8B Active Params, 70% Terminal-Bench
Poolside released Laguna S 2.1 on July 21, a 118B-parameter mixture-of-experts coding model that activates just 8B parameters per token, scores 70.2% on Terminal-Bench 2.1, and ships open weights at $0.10 per million input tokens. It does not enter our ranked leaderboard, because Poolside published no SWE-bench Verified number and we do not rank scores nobody independent has reproduced.
AI
Kimi K3 Needed 10x the Tokens to Tie Fable 5 on SWE
Fireworks AI ran Kimi K3 and Claude Fable 5 through roughly 1,030 tasks in real agent loops and found them level on software engineering, 92.4% to 92.6%. But K3 spent about 55 turns and 1.3 million tokens per task where Fable spent 21 turns and 130,000, and the headline 93% state-of-the-art figure comes from an oracle router that has to run both models first.
AI
Judge Approves Anthropic's Record $1.5B Copyright Settlement
A federal judge granted final approval on Monday to Anthropic's $1.5 billion settlement with book authors, the largest copyright recovery in US history. Roughly 500,000 works get about $3,000 each, and the plaintiffs' lawyers had their fee request cut by $86 million.
AI
Claude Fable 5 Broke a Math Conjecture Open Since 1939
Anthropic number theorist Levent Alpoge announced on July 20 that Claude Fable 5 produced a counterexample to the Jacobian Conjecture, an algebraic geometry problem open since 1939. The counterexample is a degree-7 polynomial map in three dimensions whose Jacobian determinant is a constant -2, yet three distinct points collapse onto the same output.
AI
Meta Wants to Rent Anthropic $10B of Compute
Meta is in early talks to lease Anthropic up to $10 billion of computing power over two years, which would turn Meta's ad-funded data centers into a cloud business and make it a landlord to a direct model rival.
AI
Kimi K3 Is Third Best at Coding. It Is Also Not Open.
Kimi K3 scores 93.4% on SWE-bench Verified in vals.ai's independent test, third best of 72 systems and statistically tied with Claude Fable 5 at a third of the price. But the model Moonshot calls the largest open model ever still has no downloadable weights, and its own Terminal-Bench claim came in 3.3 points high when somebody else ran it.
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.