Anthropic Talked Slowdown. Its Actions Say Otherwise
A next-gen Claude model, an IPO push, and an actual biology lab. Put the three together and Anthropic's message isn't "stop AI." It's build stronger models while scaling up verification and control at the same pace.
A new model, one week after calling for a slowdown?
On the surface it looks like a contradiction. Anthropic CEO Dario Amodei argued on September 12 that the pace of AI capability improvement needs to slow down. Then on September 19, Reuters reported that Anthropic is weighing the timing of a new model release to counter OpenAI's GPT-6 Astra.
There's a distinction worth making here. Arguing for managing the pace of capability gains is not the same as arguing for halting commercial model releases. Anthropic is evaluating the safety of its next-generation model while also working out how not to lose ground in the enterprise and developer markets.
In fact, right after the slowdown remarks, Anthropic announced it would expand independent evaluation with Accenture. The two companies plan to invest at least $1 billion each over the next five years. Rather than pause model development, the approach is to put stronger models inside more external review and more safeguards.
One model swap can shake up the entire enterprise AI market
The numbers explain why Anthropic is on edge. In enterprise AI spending tracked by expense management platform Ramp, GPT-6 Astra accounted for about 13%, while Claude Fable held about 8%. On OpenRouter too, OpenAI model spending recently overtook Anthropic's. According to Reuters, this is the first time in roughly two and a half years that OpenAI has led on that metric.
13% — GPT-6 Astra's share of enterprise AI spend tracked by Ramp 8% — Claude Fable's share of enterprise AI spend tracked by Ramp $65B+ — Anthropic's annualized revenue run rate as of July $190-200B — Reuters-reported Anthropic revenue projection for 2028
Annualized run rate converts a given moment's revenue pace into a one-year figure; it differs from confirmed annual revenue.
Still, it's too early to say Anthropic has fallen behind. According to Reuters, Anthropic's annualized revenue run rate topped $65 billion in July, ahead of OpenAI's roughly $40 billion. That's a sharp jump from about $9 billion at the end of 2025, in just seven months.
The key point is that the lead isn't fixed. Enterprise customers weigh not just model accuracy but price, security, data retention policy, coding performance, and the cost of integrating with existing systems. When a new generation of models arrives, budget allocation can shift quickly.
So why did Anthropic build an actual lab?
A second Reuters report from the same week is more striking. Anthropic is running a wet lab in the San Francisco Bay Area that conducts real biology experiments. Eric Kauderer-Abrams, the company's head of life sciences, confirmed it.
A wet lab isn't a space that only simulates things inside a computer. It's a place where researchers handle cells, proteins, reagents, and lab equipment directly. Anthropic is using it to verify Claude's computed results through actual experiments, and is studying, longer term, how AI might operate lab equipment and robots under limited human oversight.
The company has already launched Claude Science and released a Model Hardware Standard that lets AI control physical devices like microscopes, liquid-handling equipment, and robotic arms. Anthropic also said Claude recently optimized more than 30 biomolecule models in about four weeks, making them roughly four times faster on average.
The boundaries are clear, though. Anthropic says the facility isn't dedicated to any single drug-discovery program, and the company doesn't run clinical trials itself. Its acquisition of Coefficient Bio has been confirmed, but the reported deal value of about $400 million is just that, reported, not officially confirmed by Anthropic.
An AI company becoming a "lab operating system"
The significance here isn't that Claude gives good answers to doctors or researchers. The bigger shift is that AI is starting to enter the entire loop of scientific research.
| Stage | Before | What's emerging now |
|---|---|---|
| Hypothesis | Researchers read literature and narrow candidates | AI analyzes large-scale literature and data to propose candidates |
| Design | Researchers write experiment protocols | AI designs experiment plans and conditions |
| Execution | Humans operate equipment directly | AI coordinates robots and equipment through standardized interfaces |
| Verification | Humans re-analyze results | AI reads results and proposes the next experiment |
If this loop gets shorter, a single researcher can run more experiments in a day. Labs could eventually run around the clock. It's still early, and physical error rates, reproducibility, and safety issues remain significant. But if it works, AI demand expands beyond chatbot inference volume to scientific equipment, GPUs, data storage, networking, and automation hardware.
Who benefits, and who bears the cost
Cloud and AI infrastructure providers get a demand-side tailwind. If Anthropic scales up model development and scientific research at the same time, computing needs for training, inference, and lab automation are likely to grow. That doesn't automatically translate into higher profits for Amazon or Alphabet, though. Pricing and supply terms on long-term contracts, and how Anthropic splits capacity between its own procurement and multiple cloud providers, matter more than the headline demand signal.
Anthropic itself faces a two-sided picture. Revenue growth and enterprise market position are strong, but running next-generation model development, independent safety evaluation, biology research, and large compute contracts all at once could also mean burning through more cash. In an IPO, investors are likely to care as much about losses, long-term compute commitments, gross margins, and customer concentration as they do about the revenue growth rate.
The pharma and biotech industry gets a new partner and a potential rival to watch. Anthropic draws a line at not running clinical trials directly, but its customers will likely start asking how their own research data and know-how get separated from the AI provider's in-house research.
Insight Times Editorial Desk
