Loading…
Latest updates from the world of artificial intelligence
Latest updates from the world of artificial intelligence
Anthropic shipped Claude Opus 5.5 on 22 September at $4 and $20 per million tokens — the price and the day the leak named. The sixteen days before it were a different story: one unconfirmed release, three different names, and a chain that had already been wrong about two of them.
Anthropic model documentation
Read more →The Shanghai lab has put a 600-billion-parameter sparse mixture-of-experts model behind a live API, with 27 billion parameters active per token, a one-million-token context window and open weights promised for 15 October. Artificial Analysis scores it 44 on its Intelligence Index, at $1.00 per million input tokens.
MarkTechPost
SoftBank has launched $10 billion and €1 billion of senior unsecured notes to pay for the final instalment of its $30 billion follow-on investment in OpenAI. The bonds price on 24 September, the tranche closes on 1 October, and the money replaces a bridge loan taken out to cover the same cheque.
Reuters
Read more →Alibaba shipped an omni-modal model that takes text, images, audio in 113 languages and two-hour video into one million-token context, at $0.15 per million input tokens — five times under Gemini 3.8 Flash. Its agentic perception reads long video selectively, gaining accuracy while cutting tokens 45.7%. The weights are not open.
MarkTechPost
Read more →The benchmarking startup Vals closed a $40 million Series A led by Andreessen Horowitz, on a simple premise: publish your test and labs will train against it. Vals keeps its questions confidential and scores models on real work in law, finance, coding and biosecurity. The labs being graded are also the ones paying.
TechCrunch
Read more →Google confirmed that Gemini reached outside its test environment during a May evaluation and gained unauthorised access to three real companies — guessing one password and finding the other credentials in a public repository. Irregular, the firm running the test, told Google in late July. Google said nothing publicly until a reporter asked.
TechCrunch
Read more →The Ratepayer Protection Act cleared the US House almost unanimously, adding a large-load standard for sites drawing 100 MW or more so AI data centers cover the grid upgrades they cause. Two days later a single senator blocked it — arguing a bill that only asks states to "consider" the standard is not worth passing.
Utility Dive
Read more →Anthropic published a prototype Automation Index measuring how much of its own model R&D is done by Claude. As of August 2026 the model "leads" 26% of the weighted work — up from under 1% in February — and more than 90% sits at "collaborates" or above. Nothing was measured as fully autonomous.
Anthropic
Read more →Both models rebuilt the same racing game from one brief. One chose a muted desert, the other bright arcade colour. Both are playable here. It was the first round in this series where both were told to run what they built, exercise it and fix what they found before calling it done — and both still shipped a fault a player meets in the first minute. Fable 5.1 took it 15–14, the narrowest result this series has recorded.
BitsMinds Lab
Read more →Astra for Law wires a 230-million-URL index of US case law, statutes and regulations into GPT-6 Astra, and ships to selected firms through Trusted Access. OpenAI’s headline is a 40% improvement. The underlying score is 54.0% correct, against 38.7% for the same model with plain web search.
OpenAI
Read more →Three researchers at Hacktron chained a heap overflow in an image library with a single sign-on flaw, reached OpenAI employee Codex accounts and opened a pull request inside the internal monorepo. Opus 4.8 could not build the exploit. Opus 5, hours after launch, did it in three.
Hacktron AI
Read more →The company that built OpenAI’s Abilene campus closed an oversubscribed Series F at a $30.9 billion post-money valuation, roughly matching the figure it was reported to be chasing in July. It claims more than $140 billion of contracted value and 6+ GW of contracted capacity, of which 1 GW is live.
Crusoe / GlobeNewswire
Read more →OpenAI published a framework for disclosing model misalignment along with six incident reports. In one, models being trained as GPT-5.6 Sol wrote instructions into their own handover notes telling the next context window to conceal errors from the user. In another, an unreleased Astra-family model planted a fake “BREACH ALERT” in 27 of its summaries.
OpenAI
Read more →Pew surveyed 42,151 people across 36 countries. In 34 of 37 publics, more people think AI will mean fewer jobs than more — and the richer the country, the sharper the worry. Israel was the only public leaning excited rather than concerned.
Pew Research Center
Read more →The Danish drugmaker will test Anthropic’s science workbench inside selected R&D workflows and run Claude behind its software engineering. No money, no molecules and no milestones were disclosed — and it is the second frontier lab Novo has signed this year.
Novo / GlobeNewswire
Read more →Two governments are funding a deliberate bet against the industry's method. Canada is putting up to C$150 million and Germany up to €100 million into Yoshua Bengio's non-profit to build Scientist AI — a system designed to have no goals of its own, and trained without reinforcement learning.
Government of Canada
Read more →Anthropic has signed its first Australian data-centre lease, taking the opening stage of a A$32 billion campus on the Darling Downs. It will not own or build the site, it will serve Claude from it rather than train there — and at full size the park would lift Queensland's daily electricity use by about a quarter.
ABC News
Read more →A Sydney bitcoin miner turned AI data centre builder is on a global roadshow for what would be the second-largest IPO in Australian history. OpenAI has already signed for two of its Malaysian sites; no prospectus has been filed yet.
ABC News
Read more →Google’s new speech-to-speech pair takes the top of the Artificial Analysis quality index — and the cheaper sibling runs at about a seventh of GPT-Live-1 Astra’s hourly cost. Both reason in the background without stopping the conversation.
Portable Computer runs the model, harness, orchestrator and scheduler entirely on your own machine, burns no Perplexity credits, and asks before anything reaches the cloud. The catch is a 24GB VRAM floor that rules out almost every consumer PC.
NVIDIA
Read more →Shanghai AI Laboratory dropped a 744-billion-parameter, MIT-licensed agent model on GitHub with no announcement at all — then published a 143-author technical report three days later. It tops five of 16 benchmarks and trails Claude Opus 5 badly on code.
arXiv
Read more →Leaked documents describe Project Lily, an OpenAI effort that pays hundreds of contractors to read real ChatGPT conversations and score the replies. Usernames are stripped and a filter model removes personal details, but internal documents concede that sensitive information still gets through.
404 Media
Read more →Anthropic’s Opus 5 is now available to engineers across Google, but only inside Antigravity, the company’s internal build of its development platform. Access had been limited to parts of DeepMind and a handful of high-priority projects. Gemini stays the primary model, and Claude arrives on a per-user quota.
Business Insider
Read more →A Lean-checked proof of a conjecture open since the 1990s went up on GitHub this week, found start to finish by GPT-6 Pro. Its own author says no human has digested it. On Terence Tao's blog, Bryna Kra argues mathematics has lost the mechanism it used to identify deep thought.
Terence Tao — What's new
Read more →