Opus 5.5 and GPT-6 Sol land ninety minutes apart

Anthropic released Claude Opus 5.5 on Monday, the first model in a new Claude 5.5 family and, in the company's own words, its first release since it called for pacing the frontier. Anthropic says it performs at the level of Claude Fable 5.1 on most work while costing 40 percent less to run than Opus 5. Input and output are $4 and $20 per million tokens, and cache reads, which dominate agentic spending, drop 60 percent to 20 cents per million.

It is the strongest model Anthropic has recorded on its automated behavioral audit, and it ships with the biology and cybersecurity safeguards previously reserved for Fable 5.1, because it now matches Claude Mythos 5.1 in both areas. Vetted labs can apply to a Life Sciences Verification Program to work around them. One early tester completed a 680,000-line code migration in under a day. Sonnet 5.5 and Haiku 5.5 follow in the coming weeks.

OpenAI launched GPT-6 Sol and GPT-6 Luna the same afternoon, extending the GPT-6 family below Astra and halving API prices against the GPT-5.6 tiers they replace. Sol falls to $2 per million input tokens and $10 per million output; Luna falls to 10 cents and 50 cents. Both are in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, while Free and Go users get Luna in the desktop app. Neither model is in Chat yet, and Astra remains OpenAI's best model.

The two launches make an awkward pair. OpenAI's benchmark tables measure Sol against Claude Opus 5, the model Anthropic had replaced hours earlier, while Anthropic's tables cite GPT-6 Astra and GPT-5.6 Sol rather than what OpenAI shipped that afternoon. Each company's headline comparison was stale within a day of publication, which is a better guide to the pace of this market than either set of numbers.

An unreleased Pentagon review found that staff at US Central Command leaned too heavily on the artificial intelligence inside Palantir's Maven Smart System when they cleared a February target that turned out to be an elementary school in Minab, Iran. Two Tomahawk missiles killed more than 150 people, at least 123 of them children, on the opening day of the war. Bloomberg, which obtained the findings, describes a cascade of failures rather than one.

Maven surfaced the site as a recommended day-one target on the strength of stale intelligence that still catalogued it as a Revolutionary Guard facility. An analyst had logged changes to the site back in 2019, but in a system not connected to the database used to build strike packages. Central Command's civilian harm team had shrunk from ten people to one, and nobody from it reviewed the site. Palantir says it is not responsible for the underlying data, and has since added checks that re-review intelligence for disqualifying factors. It is the same failure mode as the fabricated report that nearly sent US forces to board a Chinese ship in September, without the luck.

Microsoft's Digital Crimes Unit says it has disrupted EvilTokens, a subscription cybercrime service that put AI at every stage of an attack. After tricking victims into approving a sign-in code on Microsoft's own legitimate login page, the platform's chatbot read the compromised mailbox and told criminals who the organization's money movers were, which vendor invoices to chase and whom to impersonate. It sold on Telegram for a $1,500 joining fee plus $500 a month.

Within months of launching in February 2026 it was linked to more than 12,000 compromised inboxes across over 10,000 organizations, heaviest in the United States, Canada, the United Kingdom, Australia, India and France. Microsoft seized 50 websites and disabled more than 150 supporting domains, and London's Metropolitan Police arrested two men. Investigators found that large parts of EvilTokens had themselves been vibe coded, drawing on several AI models — the barrier fell at both ends of this crime.

Toyota has begun putting its ELEY humanoid robots onto assembly lines to learn from the people already there, with workers wearing jigs shaped like the robots' fingers to teach precise hand movements. Nikkei Asia reports the group plans to invest $6.42 billion a year from 2028, placing 150,000 robots in its own automotive plants and another 250,000 at group companies making components and materials.

That is 400,000 machines arriving alongside a workforce that includes 18,000 veteran workers across 60 factories worldwide, 11 of them US plants. Executive vice president Hiroki Nakajima told Nikkei the company wants "a world where robots coexist with humans, rather than replacing them," which is a statement of intent rather than a commitment about headcount. Carmakers have run robotic arms for decades; what is new is a general-purpose machine that learns the job by watching someone do it. Hyundai's union struck in July over humanoid robots that had not yet arrived. Toyota's workers are being asked to teach theirs.

Nscale, the British neocloud spun out of crypto miner Arkon Energy two years ago, has filed to list on the New York Stock Exchange, weeks after saying a listing could come before the end of September. Its filing puts contracted backlog above $103 billion, which is future revenue from signed customer commitments rather than sales already booked. Roughly 85 percent of it sits with two customers: $43.8 billion of compute contracted to Microsoft through 2033, and a $44.6 billion agreement with Anthropic that is contingent on Nscale raising financing and that Anthropic may cancel if it misses milestones the filing itself calls stringent.

The company expects a valuation near $35 billion and is seeking to raise $3 billion. Concentration is the rule rather than the exception here: a Sona Asset Management paper cited by the Financial Times notes CoreWeave books 67 percent of revenue from Microsoft, and data center builder Applied Digital 67 percent from Oracle. Investors are being asked to price a backlog whose size depends on two counterparties staying committed.

Microsoft's latest Global AI Diffusion Report puts generative AI use at 18.8 percent of the world's working-age population in June, about one percentage point up on the first quarter, with nearly every economy growing. The United Arab Emirates leads at 73.3 percent and Singapore follows at 64.3 percent. South Korea added the most in absolute terms at 3.5 percentage points, and Japan grew fastest in relative terms.

The divide the report keeps returning to is between the Global North, at 28.8 percent, and the Global South, at 16.2 percent, and it widened again this quarter. Two caveats belong with the number. It is derived from Microsoft's own adjusted telemetry rather than a survey, and Microsoft says the next edition will count a wider set of tools and expects that to lift the figures almost everywhere, with the largest increase in China.

OpenAI has assembled an independent nine-member Advisory Group on Mathematics and Artificial Intelligence, hosted by Princeton's Institute for Advanced Study and drawn from Stanford, Harvard, Oxford and Cambridge, with several Fields Medals among its members. It will advise on how mathematical results are assessed and released. Members are unpaid, may offer and publish advice OpenAI did not ask for, and appear to get early access to the company's research.

It follows a run of OpenAI mathematical claims that turned into a reputational problem, most recently a September assertion that an internal model had closed more than 100 open problems, and before that a joint declaration by twenty-five Fields Medal winners arguing that racing to solve famous problems as a benchmark was damaging mathematics itself. OpenAI's answer is a panel carrying several of those medals. Researchers told The Verge it is a reasonable first step, but were left unclear on how the nine were chosen, how much influence they will carry, and whether a small group of prominent names can speak for a whole field.