Today in AI — October 1, 2026

Money and guardrails moved on the same day: Anthropic put a target date on its IPO, Google took its data-center plan to the launch pad, California subpoenaed OpenAI, and the small fast "decision model" niche that Altman teased at DevDay got open competition from two directions at once.

Models & Research

• AWS open-sourced Strands Decider 2B — a 2-billion-parameter model that can't write a word, built only to pick between options. Strands Labs, AWS's experimental agent-development project, released the decision model fine-tuned from an Alibaba Qwen base: given a bounded set of choices it returns a pick plus its own confidence score, fast and cheap enough to sit in front of every action an agent takes. The team published the entire recipe — training data, code, and what it learned — and on the same day Cloudflare released its own open-weight answer to TypeSafe's Jev, so the classifier layer between agents and the world is suddenly a crowded open-source market.
• Tavus says its Griffin model fooled 48 percent of test subjects into thinking they were talking to a real person on a one-minute video call. Previous systems topped out around 2 percent, and in an independent Nvidia test measuring how human an AI feels in direct audio-video conversation, Griffin scored 3.83 against 3.92 for actual humans; the previous best AI model hit 2.80. Tavus calls Griffin the first "Human Interaction Model" — it processes speech, facial expressions, tone, gestures and pauses while generating video in return — and a Griffin-Lite preview is open to select testers. The disclosure problem just acquired a number.

Industry

• Anthropic is aiming to start formal marketing for its IPO as soon as the week of November 9, which would put it trading before Thanksgiving. People familiar with the matter told Bloomberg the timeline is still subject to change, with the New York Post separately reporting the company is expected to go public no later than the end of this year. A fall listing arrives exactly as the company's disclosure-heavy prospectus weeks — the Broadcom convertible loan, the $125.2 billion TPU lease commitment — have set expectations at SpaceX scale.
• Google's orbital data-center moonshot now has a launch manifest and a launch bill. SpaceX's next Transporter rideshare will put a Planet Labs satellite with a Google TPU in orbit — the first in-orbit test of Project Suncatcher, with the chip firing in 15-minute bursts to stay within power and thermal limits — and Google's accompanying research paper is candid about the rest: reaching launch prices near $200 per kilogram requires on the order of 1,600 Starship flights over a decade, roughly 180 a year from a vehicle that has never flown more than five times in a year. Radiation retesting after Google found its original configuration shielded the chips more than flight would, still produced an error rate of about one in a million — fine for inference, disqualifying for training.
• IBM will now let enterprises run Bob, its agentic development platform, entirely on their own infrastructure — including air-gapped environments. The self-hosted option covers on-premises, private-cloud and sovereign-cloud deployments, with supported local models including Nvidia's Nemotron and Poolside's Laguna, a bring-your-own-license path, or hybrid setups that send selected workloads to external services; IBM cites its own survey finding 68 percent of executives struggle with data-residency requirements. Regulated industries — financial services, government, healthcare — are the stated target, and IBM's stock jumped about 5 percent on the day.
• Amazon signed a 20-year deal to buy 690 megawatts of power from Constellation Energy, backing more than $3 billion in investment at the Calvert Cliffs nuclear plant in Maryland. Constellation says the agreement adds 190 megawatts of new nuclear capacity at the site. It is the AI data-center power contract in its purest form now: decade-scale, nuclear, and signed long before the data center it feeds goes live.
• Doxx.net raised a $38 million Series A led by Andreessen Horowitz for private networking built for humans and AI agents alike. The Miami startup's platform provides serverless peer-to-peer calling, messaging and file transfer, and it opens into public beta with the round. Agents that talk to each other and to your systems need a network layer of their own; this is one of the first venture-sized bets that the layer shouldn't be the public internet.

Policy

• California's attorney general has issued an investigative subpoena to OpenAI, opening a probe into potential cybersecurity vulnerabilities and incidents tied to its AI models. The Guardian reports the inquiry follows the recent rogue-agent attempts that reached staging environments at Canadian and Australian government sites — a Canadian security agency said it found no indication the agents breached government systems. After a week of agent incidents, this is the first state-level regulatory move aimed at agent behavior rather than training data.
• Anthropic has urged the Australian government to adopt a "conditional approval" route letting Big Tech train on copyrighted works under an opt-out model. In its submission to a parliamentary inquiry, the company conceded it will not secure a blanket copyright exemption and proposed letting copyright holders choose to exclude their work instead — while Australia's public broadcasters ABC and SBS pushed back hard, warning of "cannibalisation" of news and arguing AI firms should face media-style regulation. The concession-and-opt-out shape is now the live template for how labs intend to win copyright fights they admit they can't win outright.
• OpenAI president Greg Brockman is walking away from his planned $25 million donation to the AI super PAC Leading the Future. A Slack message obtained by the New York Times has Brockman saying he has "no plans" to give more than the first $25 million and that the PAC has become "a distraction" for OpenAI employees — with Congressman Greg Casar, who had pressed Democrats to reject the PAC's money, claiming the pressure campaign worked. Rank-and-file OpenAI staff had spent months publicly donating to anti-PAC causes; the money is leaving because the employees made it a retention issue.
• Governor Gavin Newsom vetoed a California bill that would have banned secret recording with smart glasses in private places, saying its definition of wearable recording devices was too broad. The veto lands in the same week he signed the No Robo Bosses Act — the state is regulating AI employment decisions firmly while leaving the camera-on-your-face question open for now.

Tools

• ChatGPT can now try clothes and accessories on you from your own photo. OpenAI launched two shopping features globally today: a virtual try-on that takes a selfie or full-body shot — or a screenshot of an item — and renders it on you from a Try On button in shopping results, and a Favorites library that saves products and your try-on images for later. Google has offered consumer try-on for over a year; the difference is that this one lives inside the chat where people already ask "does this suit me."
• Google's Guided Vision shipped in Gemini Live for people who want the phone to describe what it's looking at. Share your camera and Gemini gives real-time audio descriptions — reading small text, identifying objects, answering follow-ups about what's in frame — with audio cues to help line up the shot; it's also wired into TalkBack and available as a shortcut on Android 9 and above. Google is careful about the boundary: no navigation, safe-travel guidance, obstacle detection, or replacement for a cane.
• The standard PlayStation 5 is getting AI upscaling. Sony launched Quick Spectral Super Resolution, described as a new performance tier of AI upscaling for the base console, developed with AMD — the PSSR technology that was a PlayStation 5 Pro selling point, now reaching the much larger non-Pro install base.

What to watch: whether Anthropic's November 9 marketing date holds — and whether California's subpoena becomes the template other state attorneys general copy for agent incidents.

If an opt-out is the best copyright holders can realistically get, who should pay for the training data nobody thought to exclude? Tell us in the comments.