Anthropic Says It May Imminently Hit Its Own AI-Accelerates-AI Threshold, and METR Says the Acceleration Is Real
Anthropic told prospective investors on November 2 that its internal estimates "indicate that we may imminently meet" the definition of "dramatically accelerated progress attributable to AI" in its Responsible Scaling Policy, according to a revised S-1 filing quoted by TechCrunch. The company added that it has not designated any of its current models as exceeding that threshold.
Three days later, the evaluator METR published its own assessment of the same question. Its headline estimate: Anthropic's capability growth on well-specified tasks is running at about 2x the rate it held from March 2024 to October 2025. METR said it would not be surprised by a figure as low as 1.3x or as high as 3x.
What the filing actually says
TechCrunch's headline said Anthropic is "very close" to self-building AI. The filing's wording is more hedged. The risk factor says the company has in the past "not identified" a significant acceleration and may fail to identify one in the future. It also says Anthropic does not anticipate being able to safely use Claude to "fully autonomously design and build successive generations of frontier models" without meaningful human supervision, and that it is considering updates to its Responsible Scaling Policy "to keep humans in control." TechCrunch also noted that an S-1 serves as both disclosure and marketing.
What METR found
METR's engagement ran from August 12 to October 29, with 2.5 full-time staff over eleven weeks. On Anthropic's AECI benchmark, scores have improved at 1.9x the pre-October 2025 rate (95% confidence interval 1.5x to 2.8x). Anthropic's own Haiku 5.5 system card put that figure at 1.7x, and METR attributes the gap mainly to newer unreleased models it could not test.
METR guessed that perhaps 80% of the October 2026 acceleration is due to AI, with substantial uncertainty. Self-reported staff uplift reached about 12x for one internal model, a number METR called "very loose evidence." It found no convincing bottlenecks for the next 12 months, though it said its understanding changed several times during the work.
The catch
METR said Anthropic held unilateral redaction rights and used them to remove "a majority of the evidence we found most informative." Anthropic described the redactions as aimed at confidential information that could speed up other developers. METR wrote that the public "has to rely on METR's judgment much more than would be ideal." It also said Anthropic dropped a verification of how compute is split between capabilities and safety work partway through.
The two documents measure different things. The S-1 refers to a policy threshold. METR says it does not pay special attention to the thresholds in that policy and uses its own AI R&D framing. Neither says a model has crossed into self-improvement.
Anthropic has not said whether it will revise the Responsible Scaling Policy before the offering. The filing says only that it is considering one.
Sources: TechCrunch · METR