The AI industry’s favorite tension—preach caution, ship capability—just got a fresh Reuters exclusive. According to three people familiar with the matter, Anthropic is weighing whether to release a new model to push back on OpenAI’s GPT-6 Astra, which has been eating into the enterprise narrative Anthropic has owned for months. The timing lands awkwardly next to CEO Dario Amodei’s September 12 essay arguing labs should slow the pace of capability improvements for safety reasons.
Anthropic declined to comment for the Reuters story. Nothing in the reporting locks a ship date. What it does lock is a live internal debate: competitive pressure from Astra, safety evaluation of the next model, and investor focus on when profits show up as rates rise.
Astra’s enterprise chart is the irritant
OpenAI dropped GPT-6 Astra on September 3 with claims around computer use, software engineering, cybersecurity, and professional workflows. Reuters says the model’s enterprise response has some prospective Anthropic IPO investors asking whether OpenAI could start peeling share from the company long cast as the enterprise AI leader.
Expense-platform Ramp’s latest tracked slice put Astra at roughly 13% of enterprise AI spending versus about 8% for Anthropic’s Claude Fable. On OpenRouter, which routes developer traffic across models, OpenAI reportedly led Anthropic in weekly spend for the first time in more than two and a half years. Those are short windows—and investors who already plan to ride both IPOs told Reuters they still see Anthropic’s enterprise incumbency as sticky—but the directional signal is loud enough to force a product conversation.
Slowdown essay, then a model debate
Amodei’s 3,800-word piece painted a grim picture of agent swarms outrunning human control and called for a slower cadence of capability releases. Sam Altman and Elon Musk publicly backed the thrust of that message. Reuters’ sources frame Anthropic’s potential launch discussion as happening in that rhetorical shadow: a company arguing the industry should decelerate while also deciding whether competitive reality demands another model drop before—or around—a public offering.
One person said safety evaluation of the next model is part of the release deliberation. Others said the spend debate is also about balancing new-model investment against profitability optics as higher rates make investors less patient with pure growth theater.
Revenue lead still huge; open weights still the long war
Anthropic’s annualized revenue run rate topped about $65 billion by the end of July in prior Reuters reporting, up from roughly $9 billion at the end of 2025, with a previously reported 2028 projection near $190–200 billion. OpenAI’s annualized run rate passed $40 billion in July. That gap is why some investors shrug at a single Astra week.
The harder structural challenge, people told Reuters, is open-source and open-weight competition—especially from China—that can slash token costs and let enterprises build more of their stack in-house. Meta, described as among Anthropic’s largest customers, is looking to reduce Anthropic usage as it grows internal capability, according to people familiar with the matter. Meta did not immediately comment.
IPO calendar: maybe after midterms
OpenAI’s Altman already said the company won’t list in 2026, citing safety timing. Anthropic’s offering has already slipped from earlier plans; Reuters previously said marketing was expected mid-October at earliest. Two people now say Anthropic could push its IPO to after the November U.S. midterm elections, without expecting the vote itself to reshape the deal. A pre-IPO model release would be classic market-prep theater: prove you can still answer a rival’s generation before you ask public markets for a safety-first premium.
Geeknewz take
This isn’t “Anthropic abandoned safety.” It’s that enterprise share charts don’t wait for essay consensus. If Claude’s brand is third-party scrutiny and careful releases, the next model has to show up without looking like panic shipping. If it doesn’t show up and Astra keeps the Ramp lead, the IPO roadshow inherits a different question: is safety-first still winning deals, or only winning speeches? Watch the eval language as closely as the launch rumor—because the company that just put Accenture inside the building as an embedded evaluator can’t afford a “we rushed it” footnote next to a prospectus.
Source: Reuters — Anthropic considers releasing new AI model ahead of IPO, sources say (Echo Wang, Krystal Hu, Milana Vinn, Sep 19, 2026).