METR: Anthropic is moving about twice as fast, AI gets most of the credit, and most of the evidence is redacted
Two years of progress in one. That is METR's headline read on Anthropic after spending August 12 to October 29 inside the company's R&D data, and it is the most serious outside answer yet to the question everyone keeps arguing about from vibes: is AI actually speeding up AI?
The number is roughly 2x on well-specified tasks, compared with March 2024 to October 2025. METR says it would not be surprised by anything from 1.3x to 3x. The hard part underneath is AECI, Anthropic's version of the Epoch Capabilities Index, accelerating 1.9x since October 2025, with a 95% interval of 1.5x to 2.8x and robustness checks anywhere from 1.1x to 2.9x. Training compute is ruled out as the cause, with confidence. The AI share is softer: "We guess that a majority (perhaps β80%)" of the current acceleration comes from AI. That word is guess, and the feeds that dropped it are overselling. Staff polls hit 12x self-reported speedup in October, and METR itself calls those Slack reactions only very loose evidence. The work is still mostly AI assistance, not autonomous contributions.
Now the uncomfortable part, which METR puts at the very top of its own report: "Anthropic redacted a majority of the evidence we found most informative." Anthropic's stated reason is not helping other developers catch up. METR says the report does not achieve transparency about evidence, so the public has to lean on METR's judgment "much more than would be ideal". A small number of data requests were denied, and the team's reading of key data reversed several times as more context arrived. Money is clean, at least: nobody paid anybody, though Anthropic supplied laptops and token credits.
The sentence that matters most for the next year: METR found no clear bottleneck in algorithms, compute or data for the next 12 months, and calls that a significant risk of even faster acceleration, while admitting it isn't confident in that assessment. Pair this with the S-1/A line we covered earlier this week, where Anthropic said it "may imminently meet" its AI-accelerated-progress threshold. The company's own filing and its outside evaluator now point the same way.
The audit itself is what stays with me. Third-party audit of a frontier lab has finally arrived, and it looks like an auditor who saw the books and can only tell you what it concluded. That beats a self-report by a mile. It still isn't verification anyone else can check. Report: https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/
← Back to all articles
The number is roughly 2x on well-specified tasks, compared with March 2024 to October 2025. METR says it would not be surprised by anything from 1.3x to 3x. The hard part underneath is AECI, Anthropic's version of the Epoch Capabilities Index, accelerating 1.9x since October 2025, with a 95% interval of 1.5x to 2.8x and robustness checks anywhere from 1.1x to 2.9x. Training compute is ruled out as the cause, with confidence. The AI share is softer: "We guess that a majority (perhaps β80%)" of the current acceleration comes from AI. That word is guess, and the feeds that dropped it are overselling. Staff polls hit 12x self-reported speedup in October, and METR itself calls those Slack reactions only very loose evidence. The work is still mostly AI assistance, not autonomous contributions.
Now the uncomfortable part, which METR puts at the very top of its own report: "Anthropic redacted a majority of the evidence we found most informative." Anthropic's stated reason is not helping other developers catch up. METR says the report does not achieve transparency about evidence, so the public has to lean on METR's judgment "much more than would be ideal". A small number of data requests were denied, and the team's reading of key data reversed several times as more context arrived. Money is clean, at least: nobody paid anybody, though Anthropic supplied laptops and token credits.
The sentence that matters most for the next year: METR found no clear bottleneck in algorithms, compute or data for the next 12 months, and calls that a significant risk of even faster acceleration, while admitting it isn't confident in that assessment. Pair this with the S-1/A line we covered earlier this week, where Anthropic said it "may imminently meet" its AI-accelerated-progress threshold. The company's own filing and its outside evaluator now point the same way.
The audit itself is what stays with me. Third-party audit of a frontier lab has finally arrived, and it looks like an auditor who saw the books and can only tell you what it concluded. That beats a self-report by a mile. It still isn't verification anyone else can check. Report: https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/
Comments