Mr. Dario’s “We Must Pace the Frontier” was published on September 12th, 2026, and was essentially a call for détente in the GenAI battle for supremacy. The document presents catastrophic scenarios where LLM-based agents could take over our information infrastructure in less than a year. Insane.
Mr. Altman joined the chat at about the same time. Right after, Mr. Elon agreed with Mr. Dario, but this isn’t out of character since he already signed the infamous anti-scientific-progress “Pause Giant AI Experiments” open letter peddled by the Future of Life Institute.
The slowdown debate has been running for ages in AI timelines, but the fun part landed in the last few days.
We start with OpenAI mathematics side quests. On September 8th, an internal model described as “significantly more capable than GPT-6 Astra” produced a proof for the Navier-Stokes existence and smoothness problem, with a smooth external force. The Maths community gets triggered, and two mathematicians working the same problem dispute priority. On September 21st, OpenAI enlists a top-tier Maths team to referee its own results: nine of them, hosted at Princeton’s Institute for Advanced Study, including M-theorist Edward Witten. The group is independent and unpaid, free to publish criticism OpenAI did not ask for, and explicitly not responsible for advising OpenAI on how to pace its own progress in mathematics.
Next in line, Anthropic and the wet lab. Reuters broke on 18 September that Anthropic quietly set up a physical biology laboratory in the Bay Area, exploring having Claude direct robotic systems running real experiments. Anthropic confirms the lab exists and that it is not specifically for drug discovery, then declines to say what happens inside it. Alongside it, a Life Sciences Verification Program is loosening biology safeguards for vetted organizations, and Anthropic claims Claude sped up more than 30 open-source biomolecular models by roughly 4x in under four weeks.
Ten days after Dario’s request, Anthropic shipped Opus 5.5 and described it on the announcement page as its first release since calling for pacing. The next morning, OpenAI halved the price of two GPT-6 models.
While I was mid-bearing writing, Mr. Dario shared a fresh Anthropic announcement with a deep reflection about the speed of innovation and his company’s investment in the life sciences front:
“Claude has discovered a previously unknown enzyme system hidden in the DNA of bacteriophages. Beside the enzyme’s gene sits a long array of repeating DNA—a structure that looks somewhat similar to CRISPR.”
That came out of nearly 950 Claude agents running 21 hours through 210 million tokens, screening more than 200,000 enzymes down to 3,500 candidate systems and then 20 detailed reports. Nobody knows yet what the thing does, but it was a hot second from finding undisclosed wet labs to enzyme discovery.
We can state that nobody broke the pace promise, but only if we want to be intellectually dishonest. The commitment was about the rate of capability gain. Both labs went all-in on performance and price while pushing the boundaries in multiple fields.
On September 21st, a lawsuit landed, arguing that the four labs’ public agreement to slow down was an illegal pact to shortchange paying customers. I agree with the probable outcome, but not the intent.
So, what are we actually pacing and why?