Key points
- Two new Claude models since the call to slow down
- What Amodei actually asked for
- What investors should ask before the IPO
Earlier this month, Anthropic CEO Dario Amodei called on the artificial intelligence industry to slow down. "We must slow the pace at which we improve the capabilities of AI models," he wrote.
Since then, Anthropic has released two new models in six days: Opus 5.5, followed by Sonnet 5.5 on Monday.
That sounds contradictory until you look at what Amodei wants to slow. Anthropic says these releases make existing capabilities cheaper to use. Ahead of a planned IPO, I'd want to know when its pacing commitment would actually require the company to hold something back.
What Amodei wants to slow
In his essay, "We Must Pace the Frontier," Amodei wrote that "pacing does not mean halting model training or technical progress." Anthropic committed to bringing outside evaluators into the company, which we covered when the essay came out. The broader slowdown depends on cooperation from other labs and governments.
The concern is a particular kind of advance. In its Opus 5.5 announcement, Anthropic said its calls for pacing rested "in large part on our expectation" that models able to "fully automate the work of AI research itself" could be trained soon.
Anthropic's position is that these two releases don't make that leap.
Anthropic says Opus 5.5 "performs at the level of Claude Fable 5.1 on most work." Fable 5.1 was already available. Separately, the company says Opus 5.5 costs 40% less to run than Opus 5. Outside evaluators, including METR, tested it before release.
The company is more explicit about Sonnet 5.5. Its launch post says the model "doesn't advance the frontier of our models' capabilities." Token prices are unchanged from Sonnet 5, at $2 per million input tokens and $10 per million output tokens. The claimed saving comes from needing fewer tokens: Anthropic says tasks can cost up to 30% less. A cheaper Haiku 5.5 is coming soon.
On Anthropic's account, these releases make an existing level of capability more affordable. That fits its stated approach to pacing.
Cheaper models can still spread faster
But cheaper matters. Customers who can afford to run a model more often can put it to work on more tasks. Anthropic could expand usage and revenue without crossing the capability threshold Amodei wants labs to respect.
The company is hardly stepping away from competition. Its Opus post calls the release "the new leading model" and says it beats GPT-6 Astra's top FrontierCode score "for about a fifth of the cost per task." That's a claim about a particular coding benchmark, but the sales pitch is clear: better results for less money.
We wrote last week about the price war with OpenAI that followed Opus 5.5. The competition over price and adoption is continuing alongside the call for restraint.
Pacing, as Amodei describes it, concerns how fast the most capable models improve. It leaves room for those capabilities to spread much faster.
The IPO angle
Reuters said Anthropic is expanding its lineup ahead of a planned IPO, and that enterprise customers make up about 80% of its business. Reuters has reported that the company picked Nasdaq and could start marketing the deal in mid-October at the earliest. Neither Anthropic nor Nasdaq has confirmed that.
A cheaper lineup that businesses can afford to run at scale is a better story to sell on a roadshow. That's just good business. But the essay never mentions the IPO, as we noted this month, and investors will want to know how a pacing commitment could affect the growth they're paying for.
Before the IPO, I'd want Anthropic to explain what would trigger a delay, who would make that decision, and whether Anthropic would act without its competitors. Cheaper releases fit comfortably inside its current position. The harder test comes when a model promises both a capability jump and revenue the company would have to postpone.
