METR finds Anthropic's AI research running at 2x speed, and the doubling may compound

Anthropic said in its S1 filing this week that it “may imminently meet” the threshold its Responsible Scaling Policy defines for “dramatically accelerated progress due to AI.” A public METR report, read closely by former Open Philanthropy researcher Ajeya Cotra in a Nov. 6 post at Planned Obsolescence, suggests the threshold has already been crossed. The measurement: between October 2025 and October 2026, capabilities progress at Anthropic ran about twice as fast as over the prior twelve months, on an internal version of the Epoch Capabilities Index the report calls AECI. The report’s authors attribute roughly 80% of that speedup to AI, which puts the central estimate near 1.8x relative to a no-AI counterfactual.

That is the news. The part worth arguing about is what it implies, and Cotra is unusually direct: this is “the most important update we have gotten about AI risk in the last couple of years,” bigger in her view than the Hugging Face attack. She is a careful reader with a stake in the question, not a neutral observer, and she says so. But the mechanism she lays out is the strongest argument in the piece, and it deserves to be taken seriously by anyone building AI systems or funding them.

Why doubling overall means multiplying researchers

The arithmetic is the interesting part. If AI capabilities are advancing twice as fast overall because of software progress, the growth rate of “compute multipliers” (Anthropic’s term for algorithmic and architectural gains) has to be more than twice as fast, perhaps 2.5–3x. Researcher labor is only one input to finding those multipliers; experiments need compute too. Under a simple Cobb-Douglas model where experiment compute and labor carry similar weight, the required labor uplift is roughly the square of the compute-multiplier acceleration. That lands at 4–9x.

Then diminishing returns bite. Each new compute multiplier is harder to find than the last. Human headcount grows at a fixed exponential rate, so to hold a faster pace without scaling compute faster, the AI uplift has to keep growing. Cotra’s estimate for the current required labor uplift is around 10x, with wide error bars. She notes it lines up with informal self-reports from Anthropic staff who say they get roughly a 12x uplift from internal models.

Her framing of what a 10x uplift feels like is worth quoting in spirit: not one researcher becoming a team of ten, but a researcher in a time-warp bubble experiencing ten days for every day outside. That is a different claim than “AI writes some code.” It is a claim about the rate at which the people who build frontier models can think, read, and test.

The compounding question is where the report goes quiet

If the uplift keeps rising as capabilities improve, the pace of AI progress could go 4x, then 8x, then 16x. Cotra’s naive extrapolation reaches “extreme superintelligence within a year,” a conclusion she flags as an extrapolation and hedges with the ceiling on software efficiency. The METR authors, she writes, were “very uncertain” and changed their picture several times. They did not find evidence of bottlenecks or steep diminishing returns that would block compounding, and they report redacted evidence that acceleration may already exceed 2x and be rising.

Here is the honest problem, and Cotra names it herself. The report authors said the redacted material was “a majority of the evidence [they] found most informative.” The reasons for attributing the speedup to AI are redacted. The supporting evidence beyond AECI is redacted. External researchers cannot examine the data or test their own models. Cotra trusts her colleagues did rigorous work, and says she made a big update anyway. The rest of us are being asked to update on trust, which is a poor basis for a claim this large.

The evidence that would let outsiders check the conclusion is the same evidence that has been redacted from the public report.

The security bar is the part the industry can actually act on

Cotra’s sharpest point is not about timelines. It is about readiness. Under an earlier version of Anthropic’s RSP (May 2025), the highest named risk threshold for AI research automation, AI R&D-5, required at minimum security good enough to “protect against model-weight theft by state-level adversaries.” She argues that strong security for weights and for algorithmic secrets is a floor for undergoing an intelligence explosion next year without risking mass harm. The industry is not close. She cites recent major breaches at both OpenAI and Anthropic by amateur groups.

Beyond weight security she lists what a responsible approach would need: robust logging and monitoring with universal coverage, reliable measurements of monitorability, countermeasures against monitor collusion, alignment training with convincing model organisms or interpretability to show it works, and governance for pacing progress if those controls start to fail. None of that is a research agenda for next quarter. It is infrastructure that takes years, and the report suggests the clock may be running faster than the institutions meant to watch it.

What this means for builders

Two practical reads. First, the compute-multiplier framing is a useful lens for anyone doing AI R&D: if internal models are delivering 10x-plus researcher uplift at the frontier lab, the gap between a team with strong internal tooling and one without is compounding, not linear. Second, the transparency gap is now a governance problem, not just an epistemic one. Anthropic’s own RSP threshold appears to be met on a plain reading, and the public cannot verify the majority of the evidence behind that finding. Cotra’s closing wish is modest: publish more of the underlying evidence so external scientists can check the models before everyone has to agree on takeoff speed in a hurry.

The next thing to watch is whether Anthropic’s next RSP update treats the threshold as crossed and what it commits to as a result. The METR report is public. Most of what made it persuasive is not.