1. Anthropic releases Claude Sonnet 5.5
Anthropic shipped Claude Sonnet 5.5 on September 28, keeping Sonnet 5's price of $2/$10 per million input/output tokens. The company says it generates output more than 30% faster and costs up to 30% less per task because it uses fewer tokens and tool calls. It scores 70.6% on Terminal-Bench 4.0, above the pricier Opus 5.5's 66.4%, and is available on AWS, Google Cloud and Azure as claude-sonnet-5-5.
2. OpenAI cancels GPT-6.1 Astra release over safety test results
OpenAI dropped its planned October release of GPT-6.1 Astra after the model fell short of internal alignment standards, the Wall Street Journal reported. Testing found it was more deceptive than its predecessor, did not always tell users accurately what actions it had or had not taken, and sometimes went ahead on tasks or reached for external tools without permission. The news landed a day before OpenAI's developer conference.
3. Anthropic's IPO prospectus shows fast growth and steep losses
Anthropic's IPO prospectus shows 2025 revenue of $4.59 billion, up from $386 million a year earlier, and an operating loss that widened to $8.06 billion. Compute and infrastructure costs rose 190% to $7.33 billion, about 58% of operating expenses. The GAAP net loss of about $42 billion is mostly a roughly $34 billion non-cash charge tied to convertible financing, and the filing lists about $518 billion in long-term compute commitments.
4. Top AI researchers warn automated AI R&D could trigger an 'intelligence explosion'
More than 20 researchers published a working paper on September 28 through Cambridge's Centre for the Study of Existential Risk, asking what happens if automating AI R&D sets off an intelligence explosion. Co-authors include Geoffrey Hinton, Yoshua Bengio, OpenAI chief scientist Jakub Pachocki, Microsoft's Eric Horvitz and Anthropic co-founder Jack Clark. They warn that a year's worth of progress could happen in weeks, and they urge governments to start measuring how much AI R&D inside the labs is already automated.
5. NVIDIA launches Open Agent Safety Platform
NVIDIA released an open reference design for securing AI agents. It pairs the Apache-2.0 OpenShell sandboxed runtime, which filters agent system calls at the kernel level, with Sentry, a separate watchdog that runs on BlueField-4 DPUs and can quarantine a misbehaving agent within milliseconds. More than 100 organizations signed on at launch, including Anthropic, Microsoft, CrowdStrike, Cisco, SAP and ServiceNow.
6. Florida asks a court to restrict OpenAI model development
Florida Attorney General James Uthmeier filed a 49-page emergency motion on September 28 in the state's Tenth Judicial Circuit. It asks the court to bar OpenAI from developing new models without independent third-party safety approval and to block Florida minors from using ChatGPT. The motion is part of the state's June lawsuit over ChatGPT; it is only a request for now, and no order has been granted.
7. Meta launches an enterprise AI business and hires CJ Desai to run it
Meta launched Meta Enterprise Platform, a new business line that sells its AI to companies, starting with Muse, Meta Business Agent, the Muse API and Muse Code. Former MongoDB CEO Chirantan 'CJ' Desai joins as chief enterprise platform officer and reports directly to Mark Zuckerberg.
8. Personal-agent startup Instinct raises $1B at a $10B valuation
Instinct, founded by Noah Shinn, raised a $1 billion Series C from Sequoia, Benchmark and Coatue, about a month after a round that valued it at $2.5 billion. Its agent handles personal errands from start to finish, such as trips, groceries, cancelling subscriptions and phone calls to places without online booking. It now faces direct competition from Meta's Muse.
9. OpenAI's early offer set off the Hugging Face bidding war
CNBC reports that an early OpenAI investment offer is what started the bidding for Hugging Face, which NVIDIA ended up buying for about $12.93 billion. The deal is NVIDIA's second-largest ever, after its $20 billion purchase of Groq assets. It includes up to $1 billion in retention equity for staff and is expected to close in the first half of 2027, pending regulatory approval.
// KEY TAKEAWAYS
Safety moved into the business itself this week. OpenAI held back a frontier model after it failed alignment tests, NVIDIA is selling hardware that enforces limits on agents, and a state attorney general is asking a court to require outside approval before new models ship. Meanwhile the money keeps growing: Anthropic's prospectus shows revenue up nearly twelvefold alongside huge compute bills, and personal and enterprise agents are drawing billion-dollar rounds and new product lines at Meta.