What we're watching: who gets to check OpenAI's agent numbers
Three arguments were live this week and none cleared the second review. Here is each, why it fell short, and what would get it published.
The vendor's word stopped being enough to buy agents on
The strongest candidate. On September 10 OpenAI opened its Agents API, a hosted service that runs its coding-agent machinery for any developer, per the company's announcement. On September 9 California's governor signed two laws creating a state registry of AI auditors by 2029, per the governor's office. ARC Prize, the nonprofit behind the ARC-AGI reasoning tests, published two scores for GPT-6 Astra by September 4: 99.9% inside OpenAI's own adapter and 62.7% inside the neutral one. On September 12 Anthropic's chief executive, Dario Amodei, proposed that embedded outside evaluators get badges and the right to publish. The draft argued that the outside check arrived the same week the agents went on sale.
Why it did not clear: every document that undercut the vendor's word was OpenAI's, and the draft counted OpenAI's own system card, the safety write-up published with Astra, as an outside party. The second reviewer failed it on earned position and on three undated facts. Threshold: a second lab's safety document, or a neutral benchmark on Claude Fable 5.1 or Gemini 3.8 Flash showing the same gap, with every date stamped.
The compute risk moved to customers, states and utilities
Oracle said on September 10 that most of its new AI contracts this quarter came with prepayment or customer-supplied hardware, per CNBC. Google signed a 22-year nuclear power deal in Finland on September 9, per Fortum, the Finnish utility that signed it. The Pentagon is in talks to lend $5 billion to Fluidstack, a cloud startup, for grid components, per the Wall Street Journal on September 10. Why it did not clear: the Oracle composition comes from an earnings call and has appeared in no filing, and reports conflict on how much of the backlog Oracle itself must finance; neither the Pentagon nor Fluidstack confirmed the loan; and the argument extends our August piece, Whose Demand Is the $250 Billion Really Buying?, more than it advances it. Threshold: Oracle's quarterly filing showing the prepaid share, and a confirmed loan.
Agents are selling ahead of any neutral capability test
Meta shipped Muse, a personal agent that books, buys and emails, on September 8, and two accounting-software vendors declared their agents generally available on September 9. Harvey, the legal AI company, raised $550 million the same day, per its own announcement. Why it did not clear: no independent benchmark run this week covers all three current frontier models; the Fermat's Last Theorem proof Anthropic announced on September 4 ran on a model that predates the current set; and the 74% agent-rollback figure circulating this month is vendor-funded fieldwork from January, per Sinch, the messaging vendor that paid for it. Threshold: one neutral agentic benchmark across the frontier set, or METR, the independent evaluation nonprofit, publishing its time-horizon numbers for Astra and Fable 5.1.
A quiet week honestly stated beats a manufactured argument. The first candidate is one document away from running.