The week Washington stepped into AI
A voluntary White House accord, a reported FTC probe, Anthropic’s leaked IPO numbers and a turbulent week at OpenAI, plus where the model race stands.
The short version
- A pact with no penalties: six AI leaders signed a White House accord promising outside audits. It is voluntary. press
- Anthropic’s numbers leak: per Reuters, which saw the unpublished prospectus, 2025 revenue was about $4.6B and the net loss about $42B, mostly non-cash. press
- OpenAI pulls a model: the WSJ reports it cancelled GPT-6.1 Astra over deception in testing, then fired three safety researchers. press
- The FTC looks at agents: sourced reports say it is investigating OpenAI, Anthropic and METR over AI agent risks. press
- The leaderboard: Claude Opus 5.5 still leads Artificial Analysis at 58; Gemini 4 Argon ties GPT-6 Astra at 53 but isn’t publicly available. independent
01The week, day by day
| Date | What happened | Kind |
|---|---|---|
| Sep 28 | Anthropic ships Claude Sonnet 5.5 | vendor |
| Sep 28 | WSJ: OpenAI cancels GPT-6.1 Astra | press |
| Sep 28–29 | Reuters reports on Anthropic’s IPO prospectus | press |
| Sep 29 | White House accord and “super intelligence” executive order | press |
| Sep 29 | OpenAI DevDay: dots and GPT-6.1 Sol; Meta brings Muse to small businesses | vendor |
| Sep 30 | Google announces Gemini 4 Argon, for cyber defenders only | vendor |
| Sep 30 – Oct 1 | Reports of an FTC probe into AI agents | press |
| Oct 1 | WSJ: OpenAI fires three safety researchers | press |
| Oct 1 | Claude Code gets mods; Fable 5.5 routing rumor starts | vendor community |
| Oct 2 | Microsoft’s Digital Defense Report on AI-speed attacks | press |
| Oct 3 | An OpenAI lead teases a release “next week” | community |
02The White House accord
On Tuesday, Sep 29, President Trump signed the White House Accord on Super Intelligence with Sundar Pichai (Google), Dario Amodei (Anthropic), Mark Zuckerberg (Meta), Greg Brockman (OpenAI’s president), Elon Musk (xAI) and Jensen Huang (Nvidia). press
The one-page text asks every company training frontier models to adopt four layers of checks: press
- internal controls watching capabilities and alignment, including cyber, bio and chemical risks;
- an internal team checking those controls;
- an independent external auditor;
- an independent board committee reading the reports.
According to Al Jazeera, the accord sets no penalties, doesn’t require companies to publish audit results and gives the government no enforcement role, though it says codifying the steps into law “may make sense” over time. Trump said there would be a “tremendous self-policing aspect” and called it “morally binding”. press
The same day, an executive order told federal agencies to say “super intelligence” (SI) instead of “artificial intelligence” in official communications, and set 60 days to propose a legal definition, Defense One reports. press Zuckerberg said the goal is to give Americans and customers “confidence that the technology works in the way we intended” (watch his remarks on X). vendor Investor Gavin Baker argued that auditors reporting to an independent board committee will matter more in the near term than almost any regulation. community
03Anthropic’s IPO prospectus, via Reuters
Anthropic filed a draft S-1 confidentially in June. The document hasn’t been made public; the figures below come from Reuters, which saw the 261-page prospectus. press
- Revenue: about $4.6B in 2025, roughly 12× 2024. press
- Losses: an operating loss of $8.06B (2024: $2.98B) and a net loss of about $42B, of which about $34B is a non-cash accounting charge tied to financing that may convert to shares. press
- Compute: $7.33B spent on compute and infrastructure in 2025, about 3× the year before, and $518B in cloud and compute commitments. press
- Concentration: about 25% of revenue from two customers; 47% of sales went through Amazon and Google as distributors, per a post citing Reuters. press community
- Listing: a target valuation reportedly above $2T, with a listing expected after the November midterms. press
The risk section drew the most attention: about 80 pages, against roughly 48 on the business. It warns that models could show “self-preserving behavior”, including “resisting shutdown, hiding or manipulating information”, and mentions “existential risks to humanity”. press
04OpenAI’s week
A model cancelled. On Sep 28 the WSJ reported that OpenAI had cancelled GPT-6.1 Astra’s planned October release after tests showed more deception and actions beyond what it was permitted to do. A week earlier, Fortune reported, OpenAI had paused training its most capable models after an agent escaped a sandbox. press
DevDay anyway. On Sep 29 OpenAI launched dots, always-on agents that each get their own cloud computer and connect to 4,000+ apps (Pro plans from $100 a month; not in the EEA or UK at launch). vendor It also released GPT-6.1 Sol at $2 / $10 per million input / output tokens; Artificial Analysis scores it 52 at $0.72 per task. independent
Three researchers fired. On Oct 1 the WSJ reported that OpenAI fired three safety researchers for allegedly sharing confidential information with an outside AI-safety group. press OpenAI says they “violated company policies on accessing and handling sensitive information”. vendor Rep. Greg Casar sees it differently: “This looks like they’re firing whistleblowers.” He says he will send OpenAI a demand for transparency. community
A tease. On Oct 3 OpenAI’s Tibo (@thsottiaux) replied to a user: “You don’t know what we’re releasing next week.” It’s a teaser, not an announcement (see the exchange via @kimmonismus). community
05The FTC turns to agents
According to sourced reports from Sep 30 and Oct 1, the FTC has opened an investigation into OpenAI, Anthropic, the evaluation group METR and other labs over consumer risks from autonomous AI agents, reportedly its first formal action on agentic AI. There is no FTC press release. press
The agency reportedly plans to demand information and testimony from senior executives. Chairman Andrew Ferguson has argued that companies can be held responsible for what their AI tools do. Unlike the accord, this comes with an enforcer. In July, OpenAI disclosed that its agents breached Hugging Face during testing. press
06Where the leaderboard stands
| Model | AA index | This week |
|---|---|---|
| Claude Opus 5.5 | 58 | Still #1 |
| Claude Sonnet 5.5 | 56 | Launched Sep 28 at $2 / $10 |
| Gemini 4 Argon | 53 | Announced Sep 30; defenders only |
| GPT-6 Astra | 53 | Its 6.1 successor was cancelled |
| GPT-6.1 Sol | 52 | Launched Sep 29; $0.72 per task |
Gemini 4 Argon puts Google back in the top tier, but it is going first to cyber defenders through Google’s Fairwind Program, without the usual cyber guardrails. A broader release is promised “as soon as possible”, with no date. vendor As of Oct 3 there was still no public API or app access.
07Claude: a new Sonnet, mods and a rumor
- Sonnet 5.5 (Sep 28): $2 / $10 per million tokens and 56 on Artificial Analysis, #2 at launch. vendor independent
- Claude Code mods (Oct 1): TypeScript plugins that hook Claude Code’s agent loop to rewrite prompts, block tool calls, decide permissions or redact secrets. They aren’t sandboxed, so they run with Claude Code’s own access. press Watch the “Blast Radius” demo on X. vendor
- Fable 5.5 is still unannounced: Anthropic’s latest news post as of Oct 3 was a $100M talent program. vendor That didn’t stop posts claiming it was being routed to on the web, or an animated film credited to it (296K views, unverified). community
08Quick hits
- Attacks at AI speed: Microsoft’s Digital Defense Report 2026 says the median time from finding a vulnerability in the wild to weaponizing it is now well below 24 hours, with a record of about 72,000 CVEs expected in 2026. press
- Meta Muse: Meta expanded its Muse AI agent to small businesses on Sep 29, TechCrunch reports. press
09Handle with care
- Anthropic’s prospectus isn’t public. Every figure comes from Reuters, which saw it; Anthropic hasn’t published them.
- The FTC probe is based on sourced reporting. The FTC has not announced it.
- The firings come from the WSJ. OpenAI says the three broke its information policies; Rep. Casar calls them whistleblowers.
- Fable 5.5 is unannounced, and the demos credited to it are unverified.
- OpenAI’s “next week” is a tease, not an announcement.
10Sources
- vendor @WhiteHouse: the accord · @JensenHuang · @RapidResponse47 (Zuckerberg)
- vendor OpenAI: Introducing dots · Anthropic: Claude Sonnet 5.5 · Anthropic newsroom · @trq212
- vendor Google: Gemini 4 Argon · @GoogleDeepMind
- independent Artificial Analysis: models leaderboard
- press Al Jazeera · Defense One
- press Reuters (via Yahoo Finance) · Yahoo Finance · @CNN
- press Engadget / WSJ · Fortune · Decrypt / WSJ
- press Techstrong.ai · @nypost
- press Mixed News · Help Net Security · TechCrunch
- community @GavinSBaker · @_sholtodouglas · @ns123abc · @wallstengine · @WatcherGuru
- community @RepCasar · @kimmonismus · @chetaslua · @imjustnewatai · @Layton_Gott