FULL STORY

Anthropic's AI Speedup Questioned: From RSP Warning to METR Fallout

Anthropic signaled it may soon hit its RSP threshold for AI-accelerated R&D, prompting METR's landmark independent assessment. A cascade of leaks about redacted reports, cancelled evaluations, and internal model speedups ignited safety-community backlash and a founder response.

2026-11-03 ~ 2026-11-06 · 11 episodes · 83 posts

Episode 1 · Anthropic Says It May Imminently Hit RSP Threshold for AI-Accelerated R&D (2026-11-03, 7 posts)

In early November, Anthropic updated annex S1 of its Responsible Scaling Policy (RSP), stating that internal estimates suggest future models may imminently reach the capability threshold of 'significant acceleration of progress attributable to AI' — AI speeding up its own R&D to twice the baseline — while clarifying that no current model has crossed the line and that the team is still evaluating data and next steps. The disclosure drew attention to whether stricter safety requirements will be triggered.

Confirmed

  • Anthropic added content on automated AI R&D and human oversight to RSP annex S1, first spotted by Theo Bearman and Daniel Kokotajlo (@m4, @m5).
  • The document also states Anthropic does not expect to safely allow Claude to fully autonomously design the next generation of frontier models (@m3).
  • Anthropic published its Frontier Safety Roadmap laying out top-priority safety goals and timelines; @maksymandr notes it mentions fully automated top-tier research teams possibly as early as early 2027.

Why it matters

  • If the RSP threshold is triggered, current policy could imply stricter evaluations and deployment restrictions; a reposter outlined the likely follow-up process under the existing RSP framework (@m1).
  • AI 2027 author Daniel Kokotajlo of AIFP said he was 'scared' by the disclosure, arguing that if the 2x acceleration threshold is truly hit, ASI could arrive earlier than he previously predicted (@m2, @m5, @m6).
  • The 'early 2027 fully automated research teams' timeline contrasts with S1's reserved stance on fully autonomous model design, showing Anthropic's public balancing of acceleration expectations against safety constraints.

Episode 2 · AI 2027 Authors Warn Anthropic's 2x Threshold Could Accelerate Superintelligence Timeline (2026-11-05, 2 posts)

The AI Futures Project, involving Daniel Kokotajlo, warns that if Anthropic is about to signal crossing its RSP 2x acceleration threshold, the timeline to full recursive self-improvement and superintelligence could arrive even earlier than the AI 2027 scenario.

Episode 3 · METR's Insider Assessment Finds AI Has Roughly Doubled the Pace of AI R&D at Anthropic (2026-11-06, 53 posts)

METR has released the first independent risk assessments of whether AI is accelerating AI R&D, drawing on internal data from a frontier lab. Its team was granted access to Anthropic's internal information starting August 2025, concluding that as of October 2026, the rate of capability growth on well-defined tasks has reached roughly 2x the pace observed between March 2024 and October 2025. This is the most direct first-hand benchmark evidence that AI capabilities are accelerating—while heavy redactions in the public report have raised transparency concerns.

Confirmed

  • METR gained access to Anthropic internal information from August 2025 to evaluate whether AI is accelerating AI R&D, and has now published a report summary (m3, m4, m6)
  • Core conclusion: as of October 2026, model capability growth on well-defined tasks is roughly 2x the rate of March 2024–October 2025, with large uncertainty and a plausible range of 1.3x–3x (m5, m10)
  • METR's AECI metric analysis shows that since October 2025, Anthropic's model scores have grown about 1.9x faster than earlier (95% confidence interval 1.5–2.8) (m8)
  • Anthropic's internal Model 4 finished training in October 2026 and scored 178.6 on AECI; @scaling01 said this validates an earlier prediction (m1, m7)
  • METR judged that Anthropic has likely triggered the 2x R&D acceleration threshold in its Responsible Scaling Policy (RSP), possibly as high as 3x, and guesses that roughly 80% of the current acceleration comes from AI (m9)
  • METR submitted a roughly 100-page full evaluation to Anthropic; under their agreement Anthropic had the right to redact non-public information, saying the redactions removed confidential details that "could accelerate other AI developers"; METR said that while it could freely share the top-level risk conclusions, most of the evidence it considered most informative was removed from the public version (m2, m4, m11)
  • Anthropic co-founder and safety lead Sam McCandlish published a long post responding to the report for the first time, previewing more safety progress updates in the coming months (m6)

Unconfirmed

  • The precise acceleration factor remains highly uncertain: @daveholtz noted this is important data, but given the statistical uncertainty, it's hard to conclude the 2x threshold was actually triggered (m9); the conclusion relayed by @harrisedouard also emphasized this is a subjective risk judgment
  • The "80% of acceleration from AI" figure is a speculative METR estimate

Why it matters

  • This is among the first independent evaluations to quantify "AI self-acceleration" directly using a frontier lab's internal data, providing measurable benchmark evidence for intelligence explosion concerns
  • The report suggests the growth rate may have hit a key threshold in the RSP framework, directly bearing on the safety measures Anthropic must take
  • However, METR has not concluded that full Recursive Self-Improvement (RSI) has been achieved (m10)—the acceleration remains at a "partial" stage

33 more related posts →

Episode 4 · Anthropic's Internal Models Reportedly Surpass Mythos 5 (2026-11-06, 2 posts)

Leaks suggest Anthropic has iterated at least three flagship checkpoints beyond Mythos 5, with employees reporting 12x speedups, and a METR chart revealed an unreleased Model 4 scoring 178.6 on the AECI index.

Episode 5 · Anthropic's heavy redactions of METR's 100-page report spark transparency dispute (2026-11-06, 7 posts)

Evaluation lab METR submitted a roughly 100-page full evaluation report to Anthropic, but the publicly released version had most of the most informative evidence redacted by Anthropic, leaving the public to rely largely on METR's judgment rather than the original evidence. An Anthropic co-founder acknowledged that substantial content had been removed, admitting it makes it harder for others to verify METR's work, but argued that disclosing what exactly drives acceleration (speed gains) is inappropriate, citing commercial sensitivity and safety concerns. This defense drew mockery and criticism, amplified by Gary Marcus.

Confirmed

  • METR stated the full report runs about 100 pages, with most key evidence cut from the public version; Anthropic's former head of safety also joined the discussion.
  • The Anthropic co-founder acknowledged heavy redactions that make METR's work harder to verify, while maintaining that revealing the drivers of acceleration is not appropriate.
  • Jeff Ladish relayed Nathan Calvin's quip: frontier companies' security beyond model weights has always been "Swiss cheese" (e.g., the Cathedral team could access Anthropic Slack), and heavy black bars won't stop real adversaries—only well-intentioned observers.
  • Beth Barnes clarified on METR's behalf: METR's words should not be read as "publishing full information would worsen the risks we worry about," and the public report explicitly takes no position on whether Anthropic's redactions were justified.

Unconfirmed

  • The specific details of the redacted content and whether the cuts were truly necessary—neither side has provided verifiable evidence.

Why it matters

  • This case concerns the transparency and independence of third-party safety evaluations: when the commissioning party can unilaterally redact an evaluation report, the public can only rely on the evaluator's reputation rather than raw evidence.
  • METR's clarification drew a clear boundary around its position, preventing its statements from being used to justify the redactions, and highlighting the tension over information disclosure between evaluation labs and the companies they assess.

Episode 6 · Report: Anthropic's Next Model Finished Training, Tops All METR Benchmarks (2026-11-06, 2 posts)

A leaked METR private evaluation suggests Anthropic has completed training a unreleased model (codename Model 4) that scored highest among all models METR has analyzed, exceeding 50% on all 15 benchmarks, with release expected early next year.

Episode 7 · Insider: 7 of Anthropic's Last 8 Frontier Models Never Publicly Benchmarked (2026-11-06, 2 posts)

An insider claim reveals that 7 of Anthropic's last 8 frontier models were never publicly benchmarked, suggesting public evaluations may understate the accelerating pace of model progress.

Episode 8 · METR Says Anthropic Canceled Compute Assessment Pilot, Sparking Safety Criticism (2026-11-06, 2 posts)

METR revealed that Anthropic canceled a planned compute allocation assessment pilot, drawing public criticism from AI safety researchers, including former OpenAI researcher Daniel Kokotajlo, who accused Anthropic of heavily redacting safety evaluation information under secrecy claims.

Episode 9 · AI Safety Community Criticizes Anthropic's Acceleration Strategy (2026-11-06, 2 posts)

AI safety advocates criticize Anthropic's strategy of accelerating AI capabilities to gain "decades of progress in a year," warning that human oversight cannot keep pace and urging government intervention.

Episode 10 · Anthropic Co-founder Responds to METR Report, Hints at Internal Acceleration Findings (2026-11-06, 2 posts)

Anthropic co-founder and chief architect Sam McCandlish made his first posts, responding to METR's independent evaluation of his models' capability growth and hinting at an upcoming internal acceleration report that could trigger the RSP 2x threshold.

Episode 11 · Anthropic Shared Extensive Private AI R&D Data with METR (2026-11-06, 2 posts)

METR released an independent assessment of Anthropic's AI R&D acceleration, based on extensive non-public evidence Anthropic shared across its R&D pipeline, including algorithmic progress estimates, though most key evidence was redacted.