Invest1 publisher3 min readPublished
Anthropic weighs a new model as Ramp puts Astra at 13% of tracked enterprise spend
Reuters says the deliberation includes a safety review of the model itself and a judgment about profitability, and it comes days after Dario Amodei asked the industry to slow down building more capable systems.
The Investor · Invest desk
What happened
- Reuters reported that Anthropic is considering launching a new AI model to respond to growing momentum behind OpenAI's GPT-6 Astra, which was released on September 3.
- Ramp, a corporate expense platform, put Astra at about 13% of the enterprise AI spending it tracks, against about 8% for Anthropic's Claude Fable.
- OpenRouter said its users spent more on OpenAI models than on Anthropic models last week, the first time OpenAI had led on that measure in more than two and a half years.
- The deliberations follow Dario Amodei's call, days earlier, for the industry to slow the development of increasingly capable AI systems because of safety concerns.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- decision Shipping a model now spends training and inference money in the quarter investors are asking Anthropic to show sustainable cash flow in, so the company has to pick between answering Astra and protecting margin ahead of a listing.
- constraint The safety evaluation of the next model sits inside the same deliberation as the launch question, which bounds how early the release can happen and how the company can describe it after Amodei's slow-down call.
- contradiction Ramp shows Anthropic trailing on tracked spend while its run rate sits about $25 billion above OpenAI's, so at least one of the two measures is counting buyers the other never sees.
- exposure Open-weight models cut token costs and let companies run inference internally, shrinking the third-party spend pool that both the Ramp and OpenRouter readings measure.
Ramp is a corporate expense platform, and the two models it names add to 21 percent of the enterprise AI spend it processes, leaving about four fifths of that pool with other models [1]. The gap between the two is five points [6]. OpenRouter's figure covers what its users spent over a single week [7].
The revenue lines run the other way. Anthropic went from about $9 billion of annualized run rate at the end of 2025 to more than $65 billion by the end of July, roughly 7.2 times in seven months [8][4], which puts it about 1.6 times OpenAI's July figure and $25 billion above it [3]. Astra's tracked share is also about 1.6 times Claude Fable's, pointing the other way [2].
Those two measures do not count the same purchases. One is spend flowing through an expense platform [6]; the other is annualized revenue across contracts and API. So Anthropic can lose the expense-card cohort and still book more revenue than OpenAI. The reason to care anyway is that card-paying developers and small teams are where next year's enterprise contracts tend to start.
The binding constraint in the reporting is cash. Reuters says Anthropic is weighing how much to invest in new models against the need to improve profitability, with investors pressing on when AI companies will generate sustainable cash flow [4]. At $2 trillion, a listing would price the company at about 31 times the run rate it reported in July [11][5], and it has already delayed earlier plans and could push the offering past the November U.S. midterm elections [12].
Some of the money is going into biology instead. Anthropic has set up a wet laboratory in the San Francisco Bay Area to run physical experiments [13], launched Claude Science, acquired Coefficient Bio and is exploring using Claude to direct robotic systems that carry out lab work [15]. Eric Kauderer-Abrams, the company's head of life sciences, confirmed the laboratory and said Anthropic uses both its own facilities and external partners for biological research [14]. The company has said it wants to accelerate work on diseases that traditional pharmaceutical companies may not find financially attractive [16].
I would put more weight on the profitability sentence than on the share numbers. A model shipped to answer five points of one panel spends compute in the quarter investors want discipline in, and the safety evaluation of that model sits inside the same decision [3][4]. Two things would undercut that. If Ramp's panel skews toward smaller, faster-switching buyers, the five points say very little about the $65 billion book. If open-weight models keep cutting token costs and moving inference in house [10], third-party spend share becomes a smaller prize for either company. The release date settles it: if Anthropic's next model arrives early, the company will have to say what the safety evaluation it describes concluded [3].
What to watch
- Whether Anthropic's next model ships ahead of schedule, and what the company says the safety evaluation concluded.
- Whether Ramp's next reading widens or closes the five-point gap, and whether OpenRouter's crossover holds beyond one week.
- Whether the listing goes out after the November midterms, and at what multiple of the run rate it is priced.