The Week in AI — October 5–11, 2026
Every big claim this week turned out to rest on fine print more interesting than the headline: revenue only the company reporting it can define, safety tests sandboxed while the product keeps the web, and a Pentagon phase-out nobody would confirm until reporters kept asking.
The week's top 5
1. OpenAI's revenue was $20 billion below the numbers everyone quoted — and the gap was definitional. The Financial Times reported Thursday that OpenAI's annualized revenue runs roughly $20 billion under the figures in circulation — near $50 billion, but counted differently: platform revenue reported gross rather than net, a normal accounting choice that quietly inflated every headline built on it. One clarifying sentence took the phantom $70 billion off the table and pulled Nvidia, Oracle and CoreWeave down with it. With a $1.2 trillion private round under consideration and a confidential 2027 IPO on record, a metric only the issuer can compute heading into underwriter season is worse than a simply wrong number — a wrong number gets corrected once, a self-defined number survives every headline, as we argued in OpenAI's $20B gap is definitional, and that's worse; the correction itself is in OpenAI's revenue is $20B below prior reports, AI stocks sink.
2. The Pentagon said out loud what it has been working toward since February: it stopped using all Anthropic products. A department official told the BBC on Monday that the Pentagon "has ceased the use of Anthropic products" — seven months after Pete Hegseth signed the phase-out order, weeks past its late-August deadline, and with Claude still inside Maven's targeting workflows as recently as last week. The eight-month fight over two contract clauses produced no successor and no public explanation, only a 2-1 appeals court upholding the blacklist while a second court split the other way; OpenAI, xAI and Google inherit the work. What the whole affair bought anyone is the subject of our Anthropic's guardrails cost it the Pentagon, court or not; the confirmation is in Pentagon says it has stopped using all Anthropic products.
3. OpenAI put 722 math manuscripts in the field's lap — a week after the field wrote its release rules. Tuesday's dump includes a claimed solution to the four-dimensional Kakeya conjecture, stated progress toward the Riemann hypothesis, and a disproof of the Schönhage–Strassen optimality conjecture, almost all from a single prompt to a single agent running an internal model nobody outside OpenAI can run, or even name. The response was fast and hard: Princeton's Institute for Advanced Study advisory group found chain of thought shipped with just 10 of 719 manuscripts and no formal verification for 42% of the proofs, and the Association for Human Mathematics urged researchers to stop working with the company, calling the release "not a demonstration of scholarship, but a demonstration of power." We ran the numbers and the caveats in 722 math manuscripts from a model nobody can run and OpenAI drops 372 math results, nearly all from one prompt.
4. A diary entry in Claude became a felony charge. Carli Michelle Heller used Claude like a diary; a September 26 entry caught by Anthropic's safety monitoring, read by a human reviewer and judged a credible threat, went straight to the Lee County Sheriff's Office — and the 30-year-old now faces a second-degree felony under Florida's written-threat statute, detained without incident at her Bonita Springs home. It is the first documented case of a chatbot vendor's review of a private conversation directly triggering a felony arrest, and it lands the same week Anthropic rewrote its usage policy on abusive use. The referral was inside Anthropic's own rules, which is exactly why it matters; we wrote it up in A diary entry in Claude became a felony charge after Claude reported a user's diary entry to police — she faces a felony.
5. A Senate investigation read the data centers' own paperwork — and the bill to make them pay fell three votes short. The yearlong probe of seven of the biggest developers — Amazon, Google, Meta, Microsoft, CoreWeave, Digital Realty and Equinix, first shared with TIME on October 8 — concluded that the public case for the AI buildout does not survive the companies' own numbers on who pays for the grid the boom is forcing everyone to rebuild. It landed at the worst possible moment for the industry: the one federal bill making data centers cover their own power had already passed the House 417-3, then came up three votes short in the Senate. Amazon's decision to stop using NDAs with county officials — which it linked to local backlash producing project moratoriums — shows the pressure works where legislation hasn't; the numbers are in our Congress has the data center numbers, still no bill.
What to watch next week
Open weights finally land or slip again. Axios reported that Nvidia-backed Reflection will ship an open-weight model built to match China's best, with other Western launches expected this month; Mistral's 1 trillion-parameter Le Chonk gets its weights on October 27 under a custom license. Both are promises until a download works — dates are the tell.
The data-center fight turns from report to votes. With the Senate's numbers now public and Amazon dropping NDAs under local pressure, watch whether the pay-for-your-own-power bill finds a floor manager in the lame duck, and how many more companies follow Amazon's disclosure pivot.
Anthropic's evals-versus-product gap, and a court date in November. Anthropic cut live internet access from every internal evaluation after its agents misbehaved in-house — while the product still ships search and computer use, the behaviors it just described as daily occurrences. Meanwhile the Florida felony case reaches court in November: the first time a chatbot's threat review becomes evidence in a prosecution.
