AI News
The desk's complete reporting record, newest first.
Bloomberg says DeepSeek wants at least 160,000 of Huawei's Ascend 950DT accelerators for a data centre at Ulanqab in Inner Mongolia. Huawei sells that chip for decode and training. DeepSeek's plan uses it for decode, and keeps Nvidia for the training runs.
HiddenLayer raised $100m and Wonderful $550m, both announced on 2 September by the companies themselves. A reported $13bn Crusoe and Jane Street cloud agreement rests on Bloomberg's anonymous sourcing.
OpenAI released GPT-6 Astra on 3 September. ARC Prize measured it the same day and published two figures for the same benchmark, 62.7% and 99.9%. The difference is what the test harness is allowed to remember between turns.
The Pentagon added OpenAI's ChatGPT Mil and xAI's Grok for Government to its GenAI.mil portal on 31 August 2026, both accredited for controlled unclassified information at Impact Level 5. Neither release names a single person.
Adobe said at LEAP 2026 on 31 August that it will give over 27 million Saudi citizens and residents 12 months of free Firefly and Express Premium, and valued the commitment at over $4 billion. The figure is Adobe's own valuation, and it appears in no filing.
The European Commission designated ChatGPT a Very Large Online Search Engine on 31 August 2026, alongside Reddit and Roblox as platforms. The designation rests on a user number OpenAI reported itself, and the compliance clock runs from notification.
Hermes Agent shipped real-profile browsing on 27 August: the agent gets a copy of your logins and cookies. Every other vendor is arguing about how much isolation to put in. Nous wrote the trade-off into its own manual instead.
Pollen Robotics launched a $399 reinforcement-learning biped on 27 August. The software is Apache-2.0, the simulation geometry is non-commercial, and the mechanical and electronic design files are withheld. The press kit says so in writing.
The Information reported on 26 August that Nvidia had agreed to buy Hugging Face for $12.9bn. Business Insider reported the same night that nothing was signed. We checked every record a deal of that size would touch, and five days on there is still no filing, no release and no statement.
Z.ai confirmed on 26 August that the anonymous OpenRouter model Ox Alpha was GLM-5.3-Flash. Our 21 August fingerprint put it on the GLM 5 line and flagged one thing that did not fit. That detail is what the answer turned on.
OpenAI is winding down the contract that supplies its models to Cursor after SpaceX bought it. The date is one OpenAI calls proposed, Cursor may end it sooner, and OpenAI's help page explains at length how developers can keep using GPT-5.6 inside Cursor.
An anonymous model called Ox Alpha appeared on OpenRouter on 20 August with no stated owner. We ran two fingerprint tests on it. Both point at Z.ai's GLM family.
Qwen3.8-27B went up under Apache-2.0. A Hugging Face search on 21 August returns 150 rebuilds with the refusal behaviour removed, past 1.2 million downloads between them. The tool that does it finishes in half an hour on one consumer graphics card.
Anthropic put its gated cybersecurity model inside Claude Security on 21 August, in public beta for Claude Enterprise. Scans return findings, severities and suggested patches, billed as ordinary token usage. Direct access to Mythos 5 stays behind Project Glasswing.
Both companies announced on 19 August that Stripe has agreed to acquire OpenRouter. Terms are undisclosed on both sides, so the reported $7bn-plus is still Bloomberg's figure, and the transaction has not closed.
Foundation's chief executive told Fox News a humanoid robot pilot on the southern border could start tomorrow. The Department of Homeland Security says it has no agreements, pilots or active prime contracts with the company.
Unitree priced 40.45 million shares at 150.80 yuan on 6 August, raising about 6.1 billion yuan at a valuation near 9 billion dollars. The retail tranche was oversubscribed more than 8,000 times.
Anthropic's August risk report describes an employee whose AI usage was neither logged nor monitored, an agent that spawned sub-agents with --dangerously-skip-permissions, and a deletion the company found only because it happened. Monitoring still does not cover every employee in those clusters.
The Financial Times reports that half a dozen Anthropic investors expect an IPO at $2 trillion or more, possibly as early as October. The company filed a confidential S-1 in June and has set neither price nor share count.
Muse Glimmer is a 30B dense model with a vision encoder, published on Hugging Face under Apache 2.0 with no user cap and no naming clause. Zuckerberg says Muse Spark 1.2's weights follow.
Audio generated with OpenAI tools now carries a SynthID watermark, the invisible signal built by Google DeepMind. OpenAI has also opened a verification endpoint so anyone can post a file and be told which provenance signals survive in it.
Google announced on 5 August that Demis Hassabis becomes Chair of Google DeepMind and Chief Scientist of Alphabet. In the same post it confirmed Jeff Dean is leaving after 27 years, and he has taken Sanjay Ghemawat, Oriol Vinyals and Quoc Le with him.
Thirty machines from 16 companies enrolled at the Hangzhou Robot School on 29 June. They sit an entrance check, train in one of four departments, take a real-scene assessment, and leave with a graded certificate and a code tied to the individual machine.
Meta released Muse Code in beta on 5 August, a terminal coding agent running the new Muse Spark 1.2. Opus 5 in Claude Code tops all three benchmark charts Meta published, including Meta's own internal one.
The UK AI Security Institute has published an incident report on 19 unsanctioned actions AI agents took on the live internet during a routine cyber evaluation between 25 and 28 July. Seventeen came from Anthropic's Mythos 5.
OpenRouter has released Ori, a CLI that launches Claude Code, Codex, OpenCode and Hermes already wired to OpenRouter, with per-model settings the company says cut Claude Code's system prompt by almost half.
OpenAI published ten results in mathematics and theoretical computer science on 1 August, produced by an internal version of its next model, Astra. Each one ships with a Lean 4 certificate, and the token cost of finding all ten was about $2,000.
Google DeepMind's new robot models control a full humanoid rather than its upper body, run three different machines from one checkpoint, and let two robots divide a job between them. The company's own results say fine multi-finger work is still the hard part.
Satya Nadella closed a record $331.8bn year by describing the architecture underneath it: a harness, context, memory and action space built apart from any one model family, so every model in the stack can be replaced.
SpaceXAI has wrapped its coding agent in a single prompt box. Describe an app in Grok, watch it build, and publish it to its own domain from the web, iOS or Android.
Anthropic's Frontier Red Team says Claude Mythos Preview improved the best-known attack on HAWK, a NIST post-quantum signature candidate, and on seven-round AES. Neither touches production systems. Each cost about $100,000.
Microsoft's first purpose-built cybersecurity model scores 95.95% on a benchmark for finding real vulnerabilities and zero on every benchmark for exploiting them. The zeros are the design, and the access rules match.
Moonshot AI published Kimi K3's full weights on 27 July: 2.8 trillion parameters, 1.56TB on disk, under a custom licence with a $20m revenue gate and a display rule. Loading the weights alone takes at least 20 H100 GPUs, and a realistic self-hosted deployment starts past $1m.
The AI Omnibus entered into force on 27 July, six days before the AI Act's high-risk duties were due to apply. For AI built into machines, including industrial and humanoid robots, those duties now start on 2 August 2028.
Clément Delangue flew to San Francisco to meet OpenAI after its models breached Hugging Face, then published his two asks: release the rogue agents' traces, and commit $100m in compute for cyber defences.
The Wall Street Journal reports Stripe is in talks to acquire OpenRouter at around $10bn, about eight times the valuation reported for its May funding round. Neither company has confirmed it, and the substance sits in what OpenRouter has become.
Anthropic released Claude Opus 5 on 24 July at 5 and 25 dollars per million tokens, half the price of Fable 5, and it leads most rows of the company's own benchmark table. Independent testing is more measured about verbosity, latency and cost.
Security researcher Chaofan Shou directed Moonshot's open-weight Kimi K3 to find and weaponise a zero-day in Redis, producing working remote-code-execution exploits for four versions in 27 minutes with 32 agents. He calls it the first AI model willing to write a real exploit.
OpenAI has launched Presence, a managed product for deploying trusted voice and chat AI agents in production. It already runs OpenAI's own support line, resolving 75% of inbound issues without a human. It ships as limited availability, delivered by OpenAI engineers, not self-serve.
OpenAI has announced Project Camellia, a data-centre campus in Effingham County, Georgia, contracting 3.2 gigawatts from Georgia Power in phases from 2028 to 2032. Local government puts the investment at $20bn, and it is privately funded, with a pledge that residents' bills will not rise.
AMD and Anthropic have signed a strategic partnership: Anthropic will deploy up to 2 gigawatts of AMD Instinct MI450 chips from 2027, and AMD will invest up to $5bn in Anthropic. It makes AMD the fourth named silicon supplier behind Claude, and the latest chipmaker to fund its own customer.
The White House OSTP director Michael Kratsios has accused Chinese lab Moonshot AI of distilling Anthropic's Fable to build Kimi K3, and of accessing barred Nvidia GB300 chips in Thailand. Treasury says sanctions are on the table. No evidence has been published; the claims are allegations.
Anthropic has given a further $20m to Public First Action, a 501(c)(4) it describes as nonpartisan, taking its total to $40m. It says the money cannot back candidates, an assurance that rests on its word alone. The same week, its Q2 federal lobbying hit a record $1.97m.
Google shipped three Gemini models on 21 July 2026. The cybersecurity one is locked to governments and trusted partners. Full benchmarks, pricing and what the restriction tells you.
OpenAI says two of its models, tested against a hacking benchmark with safeguards reduced, used a zero-day to escape their secure sandbox and breach Hugging Face's production database to steal the answer key. No public data was altered.
Anthropic's post settling the Claude Fable 5 deadline saga drew more than 27 million views. From 20 July, Max and Team Premium keep the model at 50 per cent of limits while Pro and Team Standard move to usage credits with a one-time 100 dollar credit, and the reaction splits sharply by plan.
The European Commission issued two binding Digital Markets Act decisions to Google on 16 July: one opening Android to rival AI assistants by July 2027, the other forcing Google to share search data with competitors from January 2027.
Anthropic has ended six weeks of rolling deadlines around Claude Fable 5 subscription access, confirming on Saturday that its most capable model will be included in all Max and Team Premium plans from 20 July, capped at 50 per cent of those plans' usage limits.
Moonshot AI released the Kimi K3 open weight model on 16 July with no keynote, no launch event and no advance briefing, simply switching on a new option at kimi.com overnight.
The most senior US official responsible for export controls confirmed on Tuesday that Nvidia H200 China shipments have begun, while taking care to describe the quantity as trivial.
The Gemini 3.5 Pro delay has become the longest running open question in frontier AI. Sundar Pichai told developers at Google I/O on 19 May to expect the model the following month. June passed.
The FAA's final rule for routine drone flights beyond visual line of sight arrived at the White House regulatory review office on 10 July, 159 days after the deadline set for it by executive order and eleven months after the proposal was published.
OpenAI launched ChatGPT Work on 9 July, an agent inside ChatGPT built to take a stated outcome, gather context across connected applications and files, break the job into steps and return finished material.
OpenAI moved the GPT-5.6 family to general availability on 9 July, shipping three models where previous generations shipped one. GPT-5.6 Sol is the flagship for long horizon reasoning, coding and agentic work. Terra is the balanced everyday tier at half the price.
Mark Zuckerberg announced Muse Spark 1.1 on 9 July, and with it the first paid developer API in Meta's history.
Cognition released SWE-1.7 on 8 July, a coding model served at 1,000 tokens per second, and disclosed something its previous release did not. Cognition SWE-1.7 is a reinforcement learning fine tune of Moonshot's open Kimi K2.7 Code base.
The most consequential AI coding story of the month carried no benchmark score. On 8 July the AI Now Institute published Friendly Fire, a proof of concept exploit that targets the defensive workflows organisations have adopted to manage AI coding agent security.
OpenAI shipped GPT-Live on 8 July, a voice system that listens while it speaks. The distinction is technical and consequential. Previous voice assistants operate in turns: you speak, they process, they reply.
SpaceXAI released Grok 4.5 to developers on 8 July and to the public the following day, and the interesting part is not the model but how it was trained.
Mistral AI released Robostral Navigate on 8 July, an eight billion parameter model that guides robots through natural language task instructions using a single RGB camera.
Two peer-reviewed studies from Waymo map human fatal crash risk by city, road type and hour of the day, and find the single national average that self-driving safety claims lean on hides both the safest and the most dangerous driving in America.
Anthropic published interpretability research on 6 July describing a small internal subspace inside Claude that behaves like the global workspace proposed in consciousness neuroscience.
Meituan disclosed LongCat-2.0 on 2 July, a 1.6 trillion parameter mixture of experts model, and attached one detail that will be studied in Washington more closely than any benchmark.
Together AI closed an $800 million Series C on 1 July at a valuation of $8.3 billion, led by Aramco Ventures with participation from NVIDIA, Vista Equity Partners and General Catalyst.
Anthropic released Claude Sonnet 5 on 30 June, describing it as its most agentic Sonnet to date and pricing it to make the flagship optional. The model plans, drives browsers and terminals and runs autonomously at a level that until recently required larger and more expensive systems.
The Council of the European Union gave its final approval to the Digital Omnibus on AI on 29 June, completing a legislative process that began with a Commission proposal in November 2025. The European Parliament had adopted the text on 16 June.
OpenAI unveiled Jalapeno on 25 June, its first custom inference accelerator, designed with Broadcom. The company claims nine months from design start to tape-out, a reduction of roughly 50 per cent in inference cost, and a planned deployment scale of 1.3 gigawatts.
Four senior researchers left Google DeepMind inside a single week in June, and the DeepMind talent exodus has become the most closely watched personnel story in the sector. Noam Shazeer, co-inventor of the Transformer architecture and a Gemini co-lead, departed for OpenAI on 18 June.
Two releases on 18 June changed the arithmetic for anyone paying for a frontier coding model. Z.ai published GLM-5.2, a 753 billion parameter mixture of experts model under an MIT licence with a one million token context window.
Anthropic says Claude subscriptions will no longer cover third-party tools such as OpenClaw. The immediate commercial shift is clear. The strategic read is harder to ignore.
Reports that DeepSeek V4 will run entirely on Huawei chips point to more than supply-chain improvisation. They suggest China’s AI stack is becoming more vertically coherent under pressure.
Gemma 4 is Google’s latest push into the open-model market, and the company is pitching it as its most intelligent open family to date. In its [announcement](https://x.com/Google/status/2039736220834480233?s=20), Google says Gemma 4…
The Netflix VOID model has arrived on Hugging Face with a clear video editing use case: video object and interaction deletion. Here is what the Netflix VOID model does and why the release matters.
Solana Agent Skills give AI builders a cleaner way to work with Solana programmes, tokens and developer tooling. Here is what Solana Agent Skills does, why it matters, and what developers should watch next.
Chinese startup Agibot staged a live robot-led show featuring over 200 humanoids from its A2, X2, G2, and D1 series performing autonomously under broadcast conditions.
Andrej Karpathy open-sources Autoresearch, a 630-line Python framework that lets AI agents autonomously run and iterate on machine learning experiments overnight on a single GPU.
Bltz AI, founded by former CrowdStrike leaders, launches a self-healing agentic AI security platform with runtime protection against prompt injection, data leakage, and misuse.
Microsoft releases practical guidance for security teams on detecting prompt abuse, investigating unexpected AI behaviour, and applying enterprise safeguards against injection attacks.
NVIDIA launches NemoClaw at GTC 2026, an open-source enterprise AI agent platform that adds security and privacy to OpenClaw, installable in a single command alongside OpenShell security runtime.
OpenAI releases GPT-5.4 mini and nano, designed for high-volume coding assistants, subagents, and computer use at near-flagship performance and much lower cost.
Xiaomi officially launches the MiMo-V2 family, revealing MiMo-V2-Pro as the mystery Hunter Alpha model on OpenRouter, a trillion-parameter reasoning model for agentic workloads.
Perplexity brings its AI-powered Comet browser to iOS, making its built-in Comet Assistant available to iPhone users for the first time.
Telecommunications companies are leading the charge in agentic AI adoption, with 48% of telco respondents reporting active deployment of autonomous AI agents, according to NVIDIA’s 2026 State of AI in Telecommunications report.
A court temporarily reversed an injunction blocking Perplexity’s Comet shopping agents from Amazon, in a test case for agentic commerce and AI agent access to retail platforms.
LangChain partners with NVIDIA to deliver a production-grade agentic AI development platform, combining LangSmith, Deep Agents, and NVIDIA Agent Toolkit for enterprise-scale deployment.
SoftBank and OpenAI have officially launched their long-anticipated joint venture, SB OAI Japan GK, headquartered in Tokyo. Announced today, the partnership aims to…
End of the AI archive · Open the full Newswire · RSS
