{"schema":"sanctity-signal/0.2","record_type":"signal","id":"metr-anthropic-acceleration-e076","domain":"safety","drop_id":"signal-e076","occurred_at":"2026-11-05","how_to_use":"This is a structured record of one real-world action and its competing interpretations, from Sanctity Signal. To build content from it (a reel script, a blog, a thread): open on the hook, ground the viewer with plain and observed, walk the dated trail, present the interpretations as live tension between named readings, and close on the before/after shift and what_to_watch. Hard rules: items under established are fact; items under unknown are never presented as fact; the evidence status (verified or reported) is stated, not dropped; sources are cited.","actor":"METR","action_type":"publication of an independent assessment of AI research and development acceleration inside Anthropic, naming Anthropic's redaction of the evidence METR found most informative","observed":"Thursday November 5, 2026: METR published Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, read at metr.org. The work ran August 12 to October 29, about 2.5 full-time equivalents over 11 weeks, and METR submitted the final public report on November 3. It asked 3 questions: whether internal capability growth accelerated, how much is attributable to AI, and whether bottlenecks could cap the next 12 months. Access covered internal benchmark scores and AECI, algorithmic efficiency data, headcount, training compute metrics, internal uplift surveys, code and pull-request data, inference spend, and about 15 staff interviews. METR was not allowed to sample non-public models, and a planned verification of compute allocation was cancelled midway. Anthropic held unilateral redaction rights. METR writes that Anthropic redacted a majority of the evidence we found most informative for these conclusions, and that the public has to rely on METR's judgment much more than would be ideal; Anthropic described the redactions as aimed at removing confidential information that could accelerate other AI developers. The headline figure: AECI, Anthropic's version of Epoch's capabilities index, has grown 1.9 times faster since October 2025 than before, with a 95% interval of 1.5 to 2.8, a trend of about 15 points a year rising to about 29. Robustness across link functions and benchmarks spans 1.1 to 2.9: probit 1.8, cauchit 2.8, complementary log-log 1.1, log-log 2.9. Non-Anthropic frontier models grew about 1.4 times faster. Anthropic's own Haiku 5.5 system card said 1.7. Internal Slack polls put self-reported uplift from the most capable internal model at about 12 times in October 2026, growing about 3.9 times a year, which METR calls very loose evidence. Interviews split 2 yes, 3 leaning yes, 4 unclear, 0 leaning no, 2 no. METR is confident the acceleration was not caused by faster growth in training compute, and guesses perhaps 80% of the October 2026 acceleration is due to AI, with substantial uncertainty; it says it would not be surprised if the truth were 1.3 or 3. It found no convincing bottleneck for 12 months, at low confidence, and reports signs of life for autonomous AI contributions to parts of the training pipeline but nothing highly significant yet. Claude was heavily involved in METR's own data collection and analysis. Not established: per-stage attribution, the split across compute, data and algorithms, and whether the index maps to real capability.","corrections":[],"evidence":{"status":"verified","note":null,"sources":[{"role":"primary","label":"METR: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, Nov 5, 2026","url":"https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/"},{"role":"corrob","label":"METR: Review of the Risks from automated R&D section in Anthropic's Risk Report, Nov 5, 2026","url":"https://metr.org/blog/2026-11-05-rd-section-anthropic-risk-report-august-2026-review/"},{"role":"corrob","label":"METR on X: AECI scores grew at 1.9x the earlier rate, 95% interval 1.5 to 2.8","url":"https://x.com/METR_Evals/status/2118412205662314935"},{"role":"corrob","label":"Crypto Briefing: METR says Anthropic's models are improving at roughly twice the historical pace","url":"https://cryptobriefing.com/metr-anthropic-ai-models-twice-pace/"}]},"strength":{"cost":"Medium","commitment":"High","reversibility":"Low","persistence":"Structural","rationale":"Commitment is high: a named organisation published a number, a method and its own limits under its own masthead, and cannot unpublish the judgment. Cost is medium, 11 weeks of staff time plus the cost of saying in public that the subject blinded the assessor. Reversibility is low, because the report is now the record and the redactions are part of it. Persistence is structural: this is the first independent measurement of self-acceleration, and every later one will be read against it. What the dials cannot score is how much the redactions cost the number."},"shift":{"prior":"Claims that AI is speeding up AI research came from the laboratories themselves.","action":"Thursday November 5, 2026: METR published Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, read at metr.org. The work ran August 12 to October 29, about 2.5 full-time equivalents over 11 weeks, and METR submitted the final public report on November 3. It asked 3 questions: whether internal capability growth accelerated, how much is attributable to AI, and whether bottlenecks could cap the next 12 months. Access covered internal benchmark scores and AECI, algorithmic efficiency data, headcount, training compute metrics, internal uplift surveys, code and pull-request data, inference spend, and about 15 staff interviews. METR was not allowed to sample non-public models, and a planned verification of compute allocation was cancelled midway. Anthropic held unilateral redaction rights. METR writes that Anthropic redacted a majority of the evidence we found most informative for these conclusions, and that the public has to rely on METR's judgment much more than would be ideal; Anthropic described the redactions as aimed at removing confidential information that could accelerate other AI developers. The headline figure: AECI, Anthropic's version of Epoch's capabilities index, has grown 1.9 times faster since October 2025 than before, with a 95% interval of 1.5 to 2.8, a trend of about 15 points a year rising to about 29. Robustness across link functions and benchmarks spans 1.1 to 2.9: probit 1.8, cauchit 2.8, complementary log-log 1.1, log-log 2.9. Non-Anthropic frontier models grew about 1.4 times faster. Anthropic's own Haiku 5.5 system card said 1.7. Internal Slack polls put self-reported uplift from the most capable internal model at about 12 times in October 2026, growing about 3.9 times a year, which METR calls very loose evidence. Interviews split 2 yes, 3 leaning yes, 4 unclear, 0 leaning no, 2 no. METR is confident the acceleration was not caused by faster growth in training compute, and guesses perhaps 80% of the October 2026 acceleration is due to AI, with substantial uncertainty; it says it would not be surprised if the truth were 1.3 or 3. It found no convincing bottleneck for 12 months, at low confidence, and reports signs of life for autonomous AI contributions to parts of the training pipeline but nothing highly significant yet. Claude was heavily involved in METR's own data collection and analysis. Not established: per-stage attribution, the split across compute, data and algorithms, and whether the index maps to real capability.","update":"An outside team with internal access measured it, and says it was blinded."},"inference":{"established":["Anthropic's capability index grew 1.9 times faster since October 2025, range 1.5 to 2.8.","METR guesses about 80% of that acceleration is due to AI itself.","Anthropic held sole redaction rights and used them on METR's best evidence."],"unknown":["That the measure tracks real work. METR says the mapping is unclear.","That the figure is robust. Alternative methods give 1.1 up to 2.9.","That AI now does the research. METR found only signs of life."]},"narrative":{"hook":"An outside evaluator measured Anthropic's progress at 1.9 times its old pace.","plain":"METR worked inside Anthropic for 11 weeks and put a number on the acceleration. Caveat: METR says Anthropic redacted most of the evidence it found most informative, and held that right alone.","story":["This is the first number on self-acceleration produced by anyone other than a laboratory. It is also a report on what outside measurement is currently worth: METR says the public must rely on its judgment much more than would be ideal. Anthropic asked for the assessment and controlled what could be shown."],"trail":[{"d":"AUG 12","t":"METR's assessment begins, with access to internal benchmark scores, compute metrics, code data and staff interviews."},{"d":"OCT 29","t":"The assessment ends. A planned verification of Anthropic's compute allocation had been cancelled midway, and METR was never allowed to sample a non-public model."},{"d":"NOV 2","t":"Anthropic tells the Securities and Exchange Commission it may imminently cross its own AI acceleration threshold. Edition 070's card."},{"d":"NOV 3","t":"METR submits the final public report to Anthropic."},{"d":"NOV 5","t":"METR publishes, with the 1.9 times figure, the 80% attribution to AI, and the statement that Anthropic redacted a majority of its most informative evidence."},{"d":"Reacts","r":true,"t":"Anthropic, which asked for this and redacted it; Epoch, whose index AECI adapts; the Securities and Exchange Commission, holding the threshold disclosure; OpenAI and Google, measured at about 1.4; the United Kingdom AI Security Institute; investors weighing an Anthropic listing."}],"glossary":[{"term":"AECI","def":"Anthropic's capability index, adapted from Epoch's. It converts many benchmark scores into one number meant to stand for how capable a model is."},{"term":"Acceleration","def":"Here a ratio of growth rates, not a level. 1.9 times means the index is climbing nearly twice as fast as it used to, not that models are twice as good."},{"term":"Redaction rights","def":"The subject of a report decides what the public may see. METR had no veto over Anthropic's cuts."},{"term":"Saturation","def":"When a model scores near the top of every test, the test stops measuring and small gains look large."}]},"judgment":{"prompt":"Which explanation do you believe?","interpretations":[{"id":"a","label":"AI is compounding now","text":"An outside measurement, not a laboratory claim, puts the rate near double.","abstain":false},{"id":"b","label":"The metric is saturating","text":"Scores above 50% compress. The index may flatter recent models.","abstain":false},{"id":"c","label":"Anthropic chose the frame","text":"It asked for the review, supplied the metric and cut the rest.","abstain":false},{"id":"d","label":"Catch-up after a slow period","text":"METR says the 2023 to 2025 baseline may have been unusually low.","abstain":false},{"id":"e","label":"Too early to call.","text":"The evidence cannot yet separate these.","abstain":true}],"distribution":{"status":"held back until you have made your call","calls":null}},"story_thread":null,"what_to_watch":[{"linked_interpretation":"a","direction":"strengthens","observable_event":"A second frontier laboratory grants an outside evaluator internal access and the published assessment reports acceleration above 1.5 times, by June 30, 2027.","window_end":"2027-06-30","state":"unresolved"},{"linked_interpretation":"a","direction":"weakens","observable_event":"METR, Anthropic or Epoch publishes a revision putting the acceleration below 1.3 times, or a successor assessment finds the index no longer tracks capability, by December 31, 2027.","window_end":"2027-12-31","state":"unresolved"},{"linked_interpretation":"c","direction":"strengthens","observable_event":"METR publishes a follow-up stating its access or redaction terms were unchanged or worse, or declines to repeat the assessment, by May 5, 2027.","window_end":"2027-05-05","state":"unresolved"}]}