The AI labs are sprinting for the exit of a price war they never chose

On September 22, Dario Amodei and Sam Altman publicly agreed to slow the race toward more capable AI. Then Anthropic cut the cost of running its new Claude Opus 5.5 by forty percent, and OpenAI — ninety minutes later — answered with two models at fifty percent off ^[1]. Two executives who had just agreed to stop racing were now racing each other to the bottom of a price list. The war has three authors, and none of them is Amodei or Altman. The first is DeepSeek's V4-Flash, which arrived in early August, ranked by the benchmarking firm Artificial Analysis as the world's cheapest well-known model — more than a hundred times cheaper to operate than Anthropic's own Claude Fable 5 ^[2]. The second is the customer: corporate clients balked at high usage costs, and the June account of the two labs' IPO filings attributed the cuts that followed to competition with Anthropic and open-source models ^[3]. The third is open-weight models — freely downloadable ones anyone can fine-tune — which Morgan Stanley says exert price pressure that is structural, not predatory ^[4]. And here is the inversion: while the American labs were cutting, the Chinese labs that started the fight were raising prices ^[5]. The war reads as forced, not chosen. The pincer theory has one problem: nobody small is dying. The startups the labs were supposedly crushing posted record valuations — Cursor nearly tripled to $29.3 billion in five months, Harvey reached $8 billion ^[6] — and the labs' own venture funds were among their biggest backers ^[7]. Max Altman, a venture investor, put the predation story on the record.

So, a really amazing strategy is to suck up all the capital, have the best funds invest in your company so they’re not investing in your competitors. — Max Altman

Nobody else's numbers agree. You do not salt the earth by paying for the seeds. Four months of cuts have now done what cuts do. Ramp's card data across 70,000 U.S. businesses shows Anthropic's realized token price — the average amount customers actually paid per token, after the discounts — fell from over $1.70 in mid-September to $0.83 by October, while OpenAI's sat at $0.45 ^[8]. In July the same series showed cheaper prices doing what cheaper prices usually do: spending per employee among the biggest-spending one percent of firms rose nearly fifty percent in a single month ^[9]. By October, the effect had broken. Spend stalled even as token volumes kept climbing ^[8]. Anthropic's prospectus, filed October 26 ^[10], stakes its demand story on compute scarcity instead — on the claim that demand is limited principally by how many chips the company can get its hands on ^[11]. It is a first-half thesis in a first-half file; October's card data had contradicted it before the document went out. From that hinge the two companies fork, and they fork hard. Anthropic is sprinting. It filed confidentially — the quiet, pre-public registration a company can still withdraw — on September 29, when the listing carried a valuation above $2 trillion ^[11]. The public S-1 came October 26 at $1.5 to $2 trillion ^[10]. The roadshow starts November 9 and pricing lands the week of Thanksgiving ^[12]. Some investors are already cutting expectations toward the low end, on what they describe as a perceived slowdown in revenue growth ^[12]. And the calendar is closing: markets price a 75.7 percent probability the Federal Reserve hikes again by December 9 ^[13], while the infrastructure firms Nscale, SB Energy, and Holtec have all delayed their own listings ^[12][14]. Anthropic is running alone into the narrowest part of the window. OpenAI is not running at all. It filed confidentially that same June, within eight days of Anthropic, and then left the paper in a drawer ^[3].

We have not decided on timing yet; it may be a while because there are things we want to do that are likely easier as a private company. — OpenAI

The filing stayed in the drawer as optionality.

We recently submitted a confidential S-1. — OpenAI

The caution was bought with documented tuition. When reports surfaced on June 26 that OpenAI might delay its own listing, the Kospi fell as much as nine percent and SoftBank more than twelve ^[15], and SpaceX's own twenty-two percent slide after its IPO was a stated factor in the delay ^[15]. OpenAI's CFO had privately cautioned the company was not ready to list ^[16]. So it funds the same war without the equity market: Broadcom arranged $50 billion secured against its custom chips, part of roughly $150 billion in chip-collateralized debt the sector lined up in a single October week ^[17], alongside a $122 billion private round at an $852 billion valuation, with part of Amazon's $50 billion contingent on reaching AGI or going public ^[18]. Two confidential filings, eight days apart in June ^[3]. One is being exercised under every visible deadline; the other sits in a drawer as optionality. Both are racing the arrival of the same arithmetic in a quarterly number. On October 26 you could watch both halves in a single news cycle: the S-1 out in the morning, and by evening OpenAI had cut GPT-6.1 Terra's price by fifty percent ^[19]. The exit, and the war it is outrunning, on the same day.