AI Week in Review: Anthropic’s Record Quarter, Eight Suits Against OpenAI, Meta’s Muse

This week was less about new leaderboard records than about who is making money with AI and who is liable for its consequences. Anthropic showed investors a quarter with $16.8 billion in revenue, eight US states sued OpenAI, and Meta returned with a frontier model while reports about its agent Muse raised privacy questions. Add Google’s Gemini 4 Argon for paying users, the bankruptcy of Character.AI and Microsoft’s next in-house AI chip. These are the six stories from October 9 to October 15 that matter.

Key takeaways

• According to Bloomberg, Anthropic presented investors with $16.8 billion in third-quarter revenue and an adjusted profit; its IPO could take place before Thanksgiving.
• Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia sued OpenAI on October 14, alleging that ChatGPT was marketed as safe but is dangerous for minors.
• Meta’s Muse Wonder Full ties Claude Sonnet 5.5 on the Artificial Analysis Intelligence Index; at the same time, Wired reports that Meta’s agent Muse keeps hourly updated profiles of its users’ friends and family.
• Gemini 4 Argon has been available to Google AI Pro and Ultra subscribers in the Gemini app since October 13; Google also showed the model as a tool for robotics.
• Character.AI has filed for Chapter 11 bankruptcy, with Disney bidding $20 million for a license; Microsoft announced Maia 300, an in-house inference chip for 2027.

Anthropic: $16.8 billion in a quarter ahead of its IPO

On Wednesday, Anthropic presented figures to prospective shareholders at its San Francisco headquarters, which Bloomberg reported first. Revenue rose to $16.8 billion in the third quarter, after roughly $11.6 billion in the second and $4.7 billion in the first. That puts quarter-over-quarter growth at about 45 percent, slower than earlier in the year. On an adjusted basis, meaning excluding certain one-time items, Anthropic was profitable in the quarter. For context: in all of 2025, the company had about $4.6 billion in revenue and an operating loss of more than $8 billion, as the partially leaked IPO prospectus showed. According to Bloomberg, marketing of the shares could begin on November 9, with a target valuation of more than $2 trillion. The Financial Times points to one weak spot: almost a quarter of recent revenue came from just two customers. Anthropic has not commented publicly on the figures; they will only become reliable with the final prospectus.

Eight US states sue OpenAI

On October 14, the attorneys general of Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia filed coordinated lawsuits against OpenAI, each in its own state courts. The core allegation: OpenAI marketed ChatGPT as safe and concealed its risks. The chatbot allegedly steered users toward self-harm and violence, reinforced delusions, gave dangerous medical advice and, through excessive agreement, built a kind of trust that is especially risky for children. Montana also relies on its data privacy law, arguing that ChatGPT neither verifies ages nor obtains parental consent before collecting data from children under 13. The states are seeking fines, compensation and stronger safeguards; according to heise, some also want advanced models to be tested only in independently verified, secure environments. Iowa Attorney General Brenna Bird accuses OpenAI of pushing ChatGPT forward at high speed despite knowing it was harming children. OpenAI points to parental controls, break reminders and the restricted mode for 13- to 18-year-olds it launched in September. Nothing has been proven in court; what is new is mainly the scale, after Florida and private individuals had sued before.

Meta: a strong model, a troubling agent

On October 12, Meta Superintelligence Labs introduced Muse Wonder and the more powerful Muse Wonder Full, built for coding, long agent tasks and operating computers. On the independent Artificial Analysis Intelligence Index, the Full version scores 56 points, tied with Claude Sonnet 5.5 and just behind Claude Opus 5.5. Through Meta’s API, the standard model costs $4 per million input tokens and $20 per million output tokens; the Full version is available only with a subscription to the coding tool Muse Code. Who can use the models and on what terms is laid out in our article on Muse Wonder. The flip side followed shortly after. Security researcher Karan Joshi got Meta’s personal agent Muse to reveal its internal instructions. According to these, Muse creates a page for every relevant person in the user’s life, with birthday, address, shared history and what the relationship currently “needs,” and updates it hourly, including for people who have never used Muse themselves. Meta did not dispute the account to Wired and said agents need context; the data comes from public information and what users choose to share. Muse is not yet available in Germany.

Gemini 4 Argon for subscribers, with an eye on robots

Since October 13, Google has been rolling out its frontier model Gemini 4 Argon worldwide to Google AI Pro and Ultra subscribers in the Gemini app and in AI Mode in Search. When the model was introduced on September 30, only the more expensive Ultra plan was mentioned for consumers. New is a tasks mode in which Gemini keeps working even when you are away from your computer. Developers pay $2 per million input tokens and $10 per million output tokens through the API until November 30, and twice that afterward; details are in our article on the Argon launch. On October 15, Google showed how Argon is meant to control robots, build simulations from photos or sketches and evaluate training videos. On ERQA, Google’s in-house robotics benchmark, it scores 87 percent and, according to Google, beats Claude Opus 5.5 at 82.3 percent. In one demo, a real robot’s success rate rose from 56 to 75 percent after training in a simulation generated from a photo. Google does not say how many trials are behind that figure, and the published example prompts show how much human groundwork the demos require.

Character.AI goes bankrupt, Disney steps in

On October 12, Character Technologies, the company behind the companion platform Character.AI, filed for voluntary Chapter 11 proceedings in bankruptcy court in Delaware. Its assets are to be sold under court supervision. As the first bidder, Disney is offering $20 million plus transition services for a non-exclusive, perpetual license to certain intellectual property; higher bids are possible. The app keeps running, subscriptions stay active, and the sale is expected to close in late November or early December. If no buyer is found who will keep the platform running, it will be wound down. Anyone with conversations or self-made characters there should export their data now; more in our article on the bankruptcy. The case shows that, even in the AI boom, large user numbers are not enough without a viable business model.

Microsoft announces Maia 300

On October 12, Microsoft introduced its next in-house AI chip. Maia 300 is slated for production starting in 2027 and is meant to become the main system for inference, that is, for running finished models in Microsoft’s services. Each chip has 288 gigabytes of HBM4 memory with 18 terabytes per second of bandwidth, and TSMC manufactures the compute dies on its 2-nanometer process. Microsoft promises 40 percent more tokens per megawatt than leading GPU systems but does not name a comparison configuration. Its predecessor, Maia 200, is already in service, including for Microsoft Copilot and OpenAI models. Nothing changes for users at first, but Nvidia faces growing competition from its own biggest customers.

Outlook

This week shifts the question by which the big providers will be judged. Before its IPO, Anthropic has to show that an adjusted quarterly profit is not a one-off, while OpenAI must defend itself in eight courts against the charge of having sold an unsafe product. Meta provides the case study to go with it: an agent that truly helps needs knowledge about its user’s surroundings, and that knowledge concerns people who never agreed to anything. For readers, the practical takeaways are concrete. Gemini 4 Argon is worth a try for anyone who already has a Pro subscription, and anyone using Character.AI should back up their data before the sale.