Credit: Below the Sky / Shutterstock Anthropic has formally decided that it meets its own threshold for AI dramatically speeding up its research. It based the call on “what we expect from future models and our overall rate of progress,” not on its current models, it wrote in its amended IPO filing on 10 November. The wording hardened over three filings in 15 days, as Alex Wilhelm noted in his newsletter Cautious Optimism. The AI research threshold sits in Anthropic’s Responsible Scaling Policy, the company’s own rulebook for when stronger safeguards apply.
The policy counts a “dramatic acceleration” as double the rate of progress seen without significant help from AI. It also requires that AI automating research could plausibly explain the doubling. Three filings in 15 days The first S-1, filed on 26 October, said Anthropic’s evidence on AI-driven acceleration came mostly from lagging indicators. “As a result, we may not identify a significant acceleration in our rate of development until after it has occurred,” the filing said. The first amendment followed on 2 November.
Anthropic wrote that it had “in the past not identified” such a speed-up until afterwards. Its estimates for future models showed it “may imminently meet” the threshold, it added. It had not designated any current model as exceeding it. What METR found inside Anthropic Three days later, the evaluation group METR published an independent assessment of Anthropic’s research pace.
Its staff worked on Anthropic devices from 12 August to 29 October, with what METR called “unprecedented access” to internal data. METR found that Anthropic’s models were improving on well-specified tasks at around twice the rate of March 2024 to October 2025. Its central estimate was 1.9 times, within a range of 1.5 to 2.8. “We guess that a majority (perhaps ≈80%) of the current (October 2026) acceleration is due to AI,” METR wrote. METR also found that AI cannot yet fully automate substantial parts of Anthropic’s model training pipeline.
Still, it saw “a significant risk that further capabilities progress could lead to even faster acceleration over the coming year.” The report carried caveats. The speed-up might be limited to well-specified tasks, METR said. Anthropic also held redaction rights over all non-public information in the report. “Anthropic redacted a majority of the evidence we found most informative for these conclusions,” METR wrote. Anthropic’s stated aim was to remove confidential information that could help other AI developers, according to METR.
The group said it took no payment, though Anthropic supplied free tokens. On 9 November, Anthropic released Claude Fable 5.2 alongside Mythos 5.2. Its system card said Mythos 5.2 does not double the internal rate of AI research progress. “Advance more slowly” The second amendment, a day later, made the call. Anthropic based it on its expectations “rather than on what we have observed from current models.” The policy covers progress the company can “observe or expect,” the filing noted.
Anthropic said it continues to assess the underlying data and will publish further reports. The same filing said Anthropic does not expect to safely let Claude fully build its successors. That would need “meaningful human supervision,” it said. It is weighing policy updates that would require human review of certain research decisions.
The seven co-founders addressed the pace in their letter to investors, which Anthropic posted on its IPO website for the roadshow. Building AI safely may require Anthropic, and ideally the rest of the industry, to modulate that pace, they wrote, “especially when this is driven by AI systems building themselves.” “There may be times when we choose to advance more slowly than we could in order to make our systems safer, even if that puts us at a competitive disadvantage,” the co-founders wrote. The roadshow video pitched the speed. In its section on compute, presented by co-founder Tom Brown, Anthropic said Claude Mythos got Claude running on AMD chips in three days “with basically no human intervention,” according to Simon Willison’s transcript. “We believe that this can accelerate even further as we approach recursive self-improvement,” the video says.
A call for a pause The AI Futures Project, the group behind the AI 2027 scenario, responded on 10 November. Citing METR’s report, all its authors now expect a “Superhuman AI Researcher” at some point in 2027, unless the industry deliberately slows down. They define it as an AI that matches the research taste of the best human researchers and fully automates coding. “We strongly urge Anthropic to release enough information such that external observers can make a complete assessment,” the authors wrote. They called for an industry-wide pause on AI research and development.
Human review of the kind Anthropic described could “provide too much comfort,” they added. Amodei urged the industry to pace the frontier in September. Four subscribers have since sued the labs under antitrust law over the idea. Anthropic’s IPO seeks a $1.4 trillion valuation, with trading on Nasdaq due to start on 19 November.
Get the TNW newsletter ✕ Close Get the TNW newsletter Get the most important tech news in your inbox each week.
Leave a Reply