METR's November 5 assessment, based on an 11-week engagement with internal Anthropic data, found that as of October 2026, Anthropic is improving AI capabilities on well-specified tasks at approximately twice the rate seen from March 2024 to October 2025, as measured by AECI benchmark scores growing at an estimated 1.9x the pre-October 2025 rate. METR attributes roughly 80% of that acceleration to Anthropic's use of AI in its own R&D workflows, not to increased training compute. METR found no significant bottlenecks in algorithmic progress, compute, or data expected to slow capabilities progress within 12 months. METR found Anthropic may have crossed its own self-defined AI acceleration warning threshold as early as July 2026. Anthropic's S1 filing stated the company may imminently meet its own RSP threshold for dramatically accelerated AI-driven progress.
The public report is substantially redacted. Anthropic exercised contractual redaction rights over the majority of evidence METR found most informative, including speedup evidence tied to specific parts of the model training pipeline. METR acknowledged the public must rely on its judgment much more than would be ideal, and one graph was labeled 'FAKE DATA' as a placeholder. Sam McCandlish stated the redactions were made because sharing what is driving the speedup risks handing others a roadmap to go faster, and that he accepts this makes it harder for others to check METR's work. Critics noted irony in citing competitive concerns while not meaningfully slowing development or supporting specific regulatory mechanisms. Anthropic also canceled a planned compute allocation audit; Daniel Kokotajlo called the stated reason 'a non-answer, a non-excuse,' arguing compute allocation data is foundational to any compute-based governance mechanism. METR noted that throughout the study it had to consider Anthropic's ability to pull access.
The methodology is contested. Toby Ord argued the apparent speedup may be closer to 1.4x if it reflects a one-time level shift from the Mythos pre-train rather than a sustained slope change, and that including non-Anthropic models also yields 1.4x. Buck Shlegeris responded that the public graph is not METR's main evidence and that METR considered and rejected the Mythos-jump hypothesis for redacted reasons. METR found that Anthropic's August 2026 risk report described the acceleration as 'less than a factor of 2' while the private figures imply 1.97x; METR stated 'a better description would seem to be around 2.'
Researcher self-reports at Anthropic describe AI uplift reaching 12x for some researchers in October 2026, with researchers reporting 80% reductions in time between idea and experiment. Researchers identified Opus 4.5 and Mythos Preview as key inflection points for autonomous coding and autonomous experiment running, with the main bottleneck shifting from experiment execution to generating new high-quality experiment ideas. Ryan Greenblatt estimated AI software R&D specifically is accelerating at roughly 4.5x, extrapolating from the reported 2x overall figure. METR estimated an overall 2x acceleration in capabilities implies a 4-9x uplift to labor specifically, because researcher labor is only one input alongside compute.
A Redwood Research post published November 9 questions this picture: Redwood researchers report only 2-3x empirical research speedup despite using Claude Code and autonomous research scaffolds, compared to the implied roughly 10x individual uplift in METR's Anthropic figures. The post raises possible explanations including Anthropic-specific infrastructure advantages, safety research being less AI-amenable than capabilities research, and researchers becoming bottlenecked on judgment rather than execution. The author states that if the gap is real, a genuine 10x speedup should produce a dramatic increase in published safety research outputs soon.