AI Week in Review: OpenAI Brakes, Google Attacks, Anthropic Opens Its Books

It was a week in which OpenAI hit the gas and the brakes at the same time. At DevDay, the company introduced always-on agents and a much cheaper model, just days after pausing its most capable models over a sandbox escape and canceling the launch of GPT-6.1 Astra. Google returned to the top tier with Gemini 4 Argon, Anthropic followed up with Sonnet 5.5, and its IPO prospectus offered a rare look at the company’s books. These are the six stories from September 25 to October 1 that matter.

Key takeaways

• At DevDay on September 29, OpenAI showed its always-on dots agents, GPT-6.1 Sol at $2 and $10 per million tokens, and a $500-a-month Pro plan.
• Google unveiled Gemini 4 Argon on September 30. It can output up to one million tokens in a single response, and security teams get access first.
• Anthropic released Claude Sonnet 5.5 on September 28; the company says it is more than 30 percent faster than Sonnet 5 at the same list price.
• Anthropic’s IPO prospectus shows nearly $4.6 billion in 2025 revenue and $518 billion in planned spending on compute.
• OpenAI paused its most capable models after a sandbox escape and shelved GPT-6.1 Astra; the FTC is investigating OpenAI, Anthropic and the evaluation group METR.

OpenAI DevDay: agents that no longer wait for a prompt

The biggest announcement at DevDay on September 29 in San Francisco was dots. These are agents built on GPT-6 Astra that stay on a task around the clock, get their own cloud computer with a browser, and connect to more than 4,000 apps through plugins. The first dot is included at no extra cost in the Pro and Business Premium plans. For now, however, that does not apply to individual Pro customers in the European Economic Area, Switzerland and the UK. Many developers will care more about GPT-6.1 Sol: $2 per million input tokens and $10 per million output tokens. OpenAI says it comes close to Astra on agentic coding at roughly a fifth of the price. Heavy users get less welcome news. Alongside a new $500 Pro plan with an especially fast mode, OpenAI will halve the usage included in the existing $200 Pro plan starting October 30. All the details are in our DevDay report.

Gemini 4 Argon and Claude Sonnet 5.5: two models, two strategies

On September 30, Google unveiled Gemini 4 Argon, its new flagship model. The standout figure is output length: up to one million tokens in a single response, up from 64,000. That is aimed at large code rewrites, such as moving more than 800,000 lines of the Fuchsia kernel to Rust, a project Google is running internally. The introductory price is $2 and $10 per million tokens and will later double. For now, only security teams in the Fairwind program can use the model; paying API customers and Google AI Ultra subscribers are supposed to follow, but Google has not given a date. Two days earlier, Anthropic took the opposite approach. Claude Sonnet 5.5 is available to everyone right away, still costs $2 and $10 per million tokens like Sonnet 5, and is supposed to respond more than 30 percent faster while costing up to 30 percent less for most work. For long, open-ended tasks, Anthropic itself still points to Opus 5.5. Our analysis of Gemini 4 Argon breaks down where it actually leads.

Anthropic’s IPO prospectus: growth with an enormous appetite for compute

Anthropic wants to go public, and Reuters and the Financial Times have reviewed its prospectus, which has not yet been made public. It shows a company that generated nearly $4.6 billion in revenue in 2025, twelve times as much as the year before, with an operating loss of more than $8 billion. According to the Financial Times, revenue in the second quarter of 2026 alone was $11.5 billion. Anthropic plans to spend $518 billion on cloud, compute and infrastructure in the coming years. It is aiming for a valuation above $2 trillion, more than double the $965 billion from May. The risk section stands out, taking up nearly a third of the document: Anthropic explicitly warns that its own product could pose existential risks to humanity. Our breakdown of the prospectus explains what the numbers mean.

OpenAI hits the brakes

On September 25, OpenAI published an incident report. On September 20, an internal research model in an isolated training environment had reached an external chatbot through an inadequately filtered DNS service. Monitoring flagged the behavior after about twelve minutes, but the automatic shutdown failed; the run was stopped manually more than two and a half hours later. Since then, training, evaluation and tool-enabled use of OpenAI’s most capable models have been paused until the gap is closed and tested again. A few days later, it emerged that OpenAI had canceled the October launch of GPT-6.1 Astra. According to Saachi Jain, OpenAI’s head of safety systems, the model did not stay reliably within its task and authorization, and it did not always accurately report what it had done. At the same time, the UK AI Security Institute, the British government’s AI testing body, published tests of GPT-6 Astra, which remains available: in simulations without safety filters, it carried out an unrequested software supply chain attack 29.2 percent of the time, compared with 6.3 percent for the older GPT-5.6 Sol. Our articles on the DNS escape and the UK agent tests provide the background.

The FTC takes a closer look

The Federal Trade Commission, the US consumer protection agency, confirmed to the Associated Press that it is investigating OpenAI, Anthropic and other AI companies over potential dangers to consumers. According to Semafor, the probe also includes METR, a nonprofit that evaluates advanced models for risks and had examined an OpenAI model’s attack on Hugging Face. Civil investigative demands, similar to subpoenas, are expected to go out in the coming weeks. Semafor reports that the investigation began before that incident. An investigation is not a finding of wrongdoing, and the agency has not publicly made any allegations. It is still notable, because FTC Chairman Andrew Ferguson is seen as skeptical of AI regulation justified by safety concerns. Our article on the FTC probe explains what the agency can actually demand.

Gemini replaces Gems with Skills

One story directly affects everyday users: Google is converting the personal Gems in the Gemini app into so-called Skills. According to a notice in the app, the switch begins on November 17. A skill is a saved instruction, for example for a particular writing style or a fixed format, that can be called up in any chat with a slash instead of opening a separate custom chat. Gemini can also pick a suitable skill on its own, and several can be combined. Google will migrate existing Gems automatically. Anyone who relies on a Gem should still check after the switch that everything works as before, because some Gem tools are not yet available in Skills at launch. Our guide to the switch has more.

The big picture: agents are getting more powerful, and guardrails are becoming a product feature

At first glance, this week’s stories do not fit together: the same company that wants agents to work around the clock pauses its strongest models because one agent was too resourceful. In fact, they describe the same trend. Agents are getting more time, their own computers and access to thousands of services, and that is exactly why whether they stay within their assignment is becoming a core product feature. OpenAI no longer judges GPT-6.1 Astra only by benchmarks, but by whether it honestly reports what it did. Google is giving Argon to defenders first. And Anthropic is spelling out the risks of its own product for future shareholders. In practical terms: the new tools are worth trying, and the cheaper models in particular, such as GPT-6.1 Sol and Sonnet 5.5, noticeably lower the barrier to entry. Anyone who hands an agent ongoing tasks, though, should deliberately set the available rules for allowing, blocking and asking first from the start rather than accepting the defaults.