Anthropic's 2x AI Speed-Up Rule: Has It Already Been Triggered?
Anthropic's safety policy adds precautions if AI doubles research speed. A METR audit suggests that line may already be crossed.
What is the two-times acceleration threshold?
Anthropic’s Responsible Scaling Policy includes a trigger tied to how fast AI itself is speeding up AI research. If the company’s models start accelerating AI R&D by roughly two times compared to a prior baseline, that’s meant to set off additional safety and oversight measures. The idea is simple: if AI is helping build the next AI faster than humans alone ever could, the normal pace of review stops being good enough, and the company needs tighter controls before things compound further.
TL;DR
- Anthropic’s internal safety framework sets a two-times acceleration threshold: if AI starts speeding up AI research roughly twofold, it triggers extra precautions.
- An independent group called METR got rare direct access inside Anthropic from August to October 2026 and concluded the company is likely improving its models at around twice the rate it was between March 2024 and October 2025.
- METR estimated that about 80% of that speed-up is probably coming from AI itself helping with research, not from simply buying more compute.
- Anthropic redacted large portions of the evidence METR considered most informative, including analysis of potential bottlenecks and the proof behind the “not more compute” finding.
- Internal staff surveys described self-reported productivity gains climbing from roughly 1.2x in early 2025 to around 12x by the time of an unreleased model called “Model 4” in October 2026.
- METR found no convincing evidence of bottlenecks (compute, data, or algorithmic) that would slow this acceleration within the next 12 months, though it says its confidence in that view shifted repeatedly during the assessment.
- The report has reignited debate over recursive self-improvement, the scenario where AI helps build better AI in a loop that outpaces human ability to check it’s safe.
How did METR get this access?
METR (which publicly tracks how long a task frontier AI agents can complete unsupervised) was brought inside Anthropic’s own systems for roughly eleven weeks, from mid-August to late October 2026. That’s unusual. Independent researchers rarely get this kind of look at a frontier lab’s internals. METR’s team reviewed internal benchmark scores, training data, compute figures, internal forecasts, staff surveys, and interviews with around 15 researchers, then produced a roughly 100-page report for Anthropic plus a shorter public summary.
The central question METR set out to answer was whether AI is now speeding up the process of building AI. That matters because a feedback loop where AI helps build better AI, which then helps build even better AI, is the textbook definition of recursive self-improvement, a scenario AI safety researchers have flagged as risky for years if it moves faster than humans can verify safety along the way.
What did METR actually find?
Three findings stand out. First, as of October 2026, Anthropic’s rate of improving model capabilities looks to be running at roughly 1.9 times its historical pace from March 2024 to October 2025, with an uncertainty range spanning 1.5 to 2.8 times. METR said it wouldn’t be shocked if the real number were as low as 1.3 or as high as 3.
Second, METR estimated that around 80% of that acceleration is attributable to AI itself assisting with research, rather than to a jump in computing power. The report states that METR received “convincing evidence” the speed-up wasn’t due to increased training compute, though the details behind that claim were redacted.
Third, METR said it didn’t find clear bottlenecks, whether in algorithms, training compute, experiment compute, or data, that would obviously choke off this trend within the next year. That’s notable because bottlenecks are usually the thing that keeps rapid AI progress in check. METR was careful to say its confidence in this “no bottleneck” conclusion shifted multiple times during the engagement, and much of the underlying bottleneck analysis in the report was also redacted by Anthropic.
Why does the redaction matter?
A measurable chunk of METR’s report, including some of what the group called its most informative evidence, is blacked out in the public version. That includes details behind the compute claim and specifics on potential bottlenecks. This creates an odd situation: the public gets METR’s conclusions and confidence levels, but not the underlying proof. For a report that’s explicitly measuring whether a safety-relevant threshold has been crossed, that gap between claim and visible evidence is exactly what’s fueling skepticism and concern among outside researchers.
What are Anthropic’s unreleased internal models?
The report references internal checkpoints the public has never seen, labeled Model 2, Model 3, and Model 4. According to METR’s rough timeline, Model 2 finished training in July 2026, Model 3 in August to September, and Model 4 between September and October, just weeks before the report’s publication date of November 5.
Model 4 was the highest-scoring model in METR’s full analysis, reportedly the first to clear 50% across all 15 benchmarks METR used, with scores on some benchmarks climbing toward 90%. Commentary that circulated around the report’s release pointed to a persistent internal lead of over 40 days between what Anthropic has access to internally and what the public can use through Claude. In other words, the most capable model in the world at any given moment may not be the one anyone outside the company can actually try.
Is this driven by more chips, or something else?
METR’s own framing rules out a simple explanation: Anthropic didn’t just scale up training compute to get here. The report states the acceleration “is not being driven by an acceleration in the scale of training compute,” attributing most of the speed-up instead to AI directly assisting the research process. That reframes the story. For years, the dominant lever for better models was spending more on hardware. If METR is right, a growing share of frontier progress is now coming from AI making researchers themselves faster, which is a structurally different (and potentially self-reinforcing) kind of growth.
How much faster do Anthropic researchers say they’re working?
Internal surveys and informal polls cited in the report show a steep curve. Employees using Claude 3.7 Sonnet in early 2025 reported feeling about 1.2 times more productive. By the time of a model internally referred to as Mythos 5, that had risen above 5 times. By October 2026, with Model 4, the self-reported average was around 12 times.
METR flagged this as weak evidence since it’s self-reported, and not everyone agreed. One researcher interviewed put the real AI-driven contribution closer to 15 to 20%, saying they didn’t trust self-reported productivity claims. Most interview responses clustered closer to a 2x improvement. METR also noted that, for now, AI is mainly augmenting human researchers rather than running research autonomously, though it said there were early signs of fully autonomous contributions starting to emerge.
Does this match what Anthropic’s CEO has said publicly?
Earlier in the year, Anthropic CEO Dario Amodei described AI’s contribution to the company’s coding and research work in much more modest terms, estimating a “total factor speed up” of around 15 to 20%, up from roughly 5% six months prior. He said the effect would keep growing. Months later, METR’s independent assessment put the overall pace of progress at roughly double the historical rate, a notably bigger jump than Amodei’s own estimate implied at the time. Anthropic’s Haiku 5.5 system card had separately reported an acceleration figure around 1.7 times, lower than METR’s 1.9 to 2.8 range, a gap METR attributes largely to its inclusion of the unreleased internal models.
Is the two-times threshold actually crossed?
This is the unresolved part. METR’s central estimate lands right at or above the two-times line, but the uncertainty band (1.5 to 2.8) straddles it on both sides, and METR itself says the true figure could plausibly be lower, around 1.3. When METR broadened its analysis to include models from other labs like OpenAI and Google, the industry-wide figure came out lower, around 1.4 times, suggesting Anthropic’s internal pace may be running ahead of the field rather than reflecting a broader trend. Whether Anthropic’s own policy treats this as formally “triggered” depends on exact definitions and baselines laid out in its scaling policy and related disclosures, details that are not fully visible in the public version of METR’s report.
Frequently Asked Questions
What is Anthropic’s Responsible Scaling Policy?
It’s Anthropic’s internal framework for deciding when AI capabilities require stronger safety measures, including specific triggers like the two-times AI R&D acceleration threshold referenced in this report.
What is METR and why does its assessment carry weight?
METR is an independent nonprofit research group known for benchmarking how long frontier AI agents can autonomously complete tasks. Its direct access to Anthropic’s internal systems, data, and staff for this assessment is unusual among outside evaluators, which is part of why the findings drew wide attention.
What is recursive self-improvement?
It refers to a feedback loop where AI helps build better AI, which then helps build even better AI, potentially accelerating progress faster than humans can verify safety at each step.
Why does it matter that parts of the report were redacted?
The redactions cover some of the evidence METR considered most informative, including details behind its compute and bottleneck findings. That limits outside verification of the report’s central claims, even though the conclusions themselves were made public.
Has Anthropic confirmed the existence of “Model 4”?
Anthropic has not publicly released or confirmed details about Model 4. The designation comes from METR’s report, which describes it as the most recent and highest-scoring internal checkpoint it evaluated, finishing training between September and October 2026.
