{"slug": "pace-of-ai-capability-growth-is-in-oct-2026-is-2x-of-mar-2024-to-oct-2025-period", "title": "Pace Of AI Capability Growth Is In Oct 2026 Is 2x Of Mar 2024 to Oct 2025 Period: METR", "summary": "METR, the nonprofit AI evaluator, reported that Anthropic is likely improving its models' capabilities on well-specified tasks at roughly twice the rate it did between March 2024 and October 2025, based on an eleven-week assessment conducted inside Anthropic from August 12 to October 29, 2026. METR estimates Anthropic's internal AECI capabilities index grew at 1.9x its earlier rate since October 2025 (95% confidence range 1.5x to 2.8x) and attributes about 80% of the acceleration to Anthropic's own use of AI in its research process, while finding no convincing evidence of compute, data or algorithmic bottlenecks that would slow growth over the next 12 months. Anthropic, which held unilateral redaction rights over the roughly 100-page report, removed a majority of the evidence METR found most informative, including detailed breakdowns of compute, data and algorithmic contributions to the acceleration.", "body_md": "There’s now some data that indicates that some degree of Recursive Self-Improvement might already be happening.\n\nMETR, the nonprofit AI evaluator, has published the results of an eleven-week assessment inside Anthropic examining whether AI is now accelerating its own development. The headline conclusion: as of October 2026, Anthropic is likely improving its models’ capabilities on well-specified tasks at roughly twice the rate it did between March 2024 and October 2025, with most of that acceleration attributable to Anthropic’s own use of AI in its research process.\n\nThe assessment ran from August 12 to October 29, 2026, with a team of 2.5 full-time-equivalent METR staff working inside Anthropic to study its AI R&D pipeline, from pretraining through reinforcement learning. This is distinct from the “employee-like access” evaluator program Anthropic [committed to](https://officechai.com/ai/anthropic-ceo-dario-amodei-says-ai-development-must-slow-down-to-pace-the-frontier/) as part of CEO Dario Amodei’s “pacing the frontier” proposal; METR describes this as a separate, earlier-starting engagement focused specifically on measuring acceleration.\n\n## The Three Questions\n\nMETR structured the assessment around three questions: is Anthropic’s capability growth accelerating, how much of that is due to AI itself, and are there bottlenecks that could slow things down in the near term.\n\nOn the first question, METR says it believes Anthropic’s rate of progress on well-specified tasks is about 2x what it was in the prior 18 months, though with wide uncertainty — the team says it wouldn’t be surprised by anything from 1.3x to 3x. Much of this is grounded in AECI (Anthropic’s internal version of Epoch AI’s Capabilities Index), which METR calculates grew at 1.9x its earlier rate since October 2025, with a 95% confidence range of 1.5x to 2.8x. In practical terms, METR says Anthropic appears to have made as much benchmark progress in the last year as it did in the two years before that.\n\nOn attribution, METR says it is confident the acceleration isn’t being driven by faster growth in training compute. Instead, the team’s best guess is that around 80% of the current acceleration comes from AI itself being used in Anthropic’s research process, though it notes there was no consensus among Anthropic staff on this point, and that non-AI factors likely explain much of the acceleration observed in late 2025.\n\nOn bottlenecks, the report says METR did not find convincing evidence of constraints — in compute, data, or algorithmic progress — that would meaningfully slow growth over the next 12 months, which it flags as a reason to take the risk of further acceleration seriously, while cautioning that its own view on this shifted multiple times during the assessment and may still be missing things.\n\n## What The Evidence Showed\n\nInterviews with Anthropic researchers broadly supported a roughly 2x acceleration story, though views varied widely by team. Separately, informal Slack polls asking staff to self-report productivity uplift from the latest internal model showed average reported gains reaching around 12x by October 2026 — a number METR treats cautiously, noting the polls reflect low-effort, likely unrepresentative responses. The report also notes that fully autonomous AI contributions to the most important parts of model training aren’t yet significant, though researchers described “signs of life” suggesting that could change.\n\n## Redactions Limited What The Public Can See\n\nMETR submitted a roughly 100-page full report to Anthropic, which held unilateral rights to redact non-public information before the public version was released. By METR’s own account, Anthropic redacted a majority of the evidence the team found most informative for its conclusions, including detailed breakdowns of compute, data, and algorithmic contributions to the acceleration, and more granular acceleration estimates for specific parts of the training pipeline. METR says this means the public has to rely more heavily on its subjective judgment than it would like.\n\nThe report also discloses that a planned compute-allocation verification component — checking how much compute Anthropic puts toward capabilities versus safety work — was cancelled by Anthropic partway through, before METR reached any conclusions on it. Anthropic said it wanted the engagement to stay focused on risk assessment for now and would revisit compute allocation later.\n\nThe findings arrive against a backdrop of recurring questions about how independent Anthropic’s outside evaluators really are, after [criticism of METR’s funding ties to Anthropic](https://officechai.com/ai/metrs-independence-questioned-after-x-user-highlights-financial-links-between-company-and-anthropic/) and separately, scrutiny over Anthropic [describing Accenture as an “independent evaluator”](https://officechai.com/ai/anthropic-community-noted-on-x-for-calling-accenture-an-independent-evaluator-of-its-ai-despite-their-business-relationship/) despite an existing commercial relationship between the two companies. METR’s report includes its own conflict-of-interest disclosure, noting no staff hold Anthropic equity, alongside a public checklist against an industry evaluator-transparency framework, in which it acknowledges it did not have query access to Anthropic’s unreleased models and could not share enough methodological detail for full independent review, owing to the redactions.", "url": "https://wpnews.pro/news/pace-of-ai-capability-growth-is-in-oct-2026-is-2x-of-mar-2024-to-oct-2025-period", "canonical_source": "https://officechai.com/ai/pace-of-ai-capability-growth-is-in-oct-2026-is-2x-of-mar-2024-to-oct-2025-period-metr/", "published_at": "2026-11-05 19:58:32+00:00", "updated_at": "2026-11-05 20:19:52.760308+00:00", "lang": "en", "topics": ["ai-research", "ai-safety", "artificial-intelligence", "machine-learning"], "entities": ["METR", "Anthropic", "Dario Amodei", "AECI", "Epoch AI"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/pace-of-ai-capability-growth-is-in-oct-2026-is-2x-of-mar-2024-to-oct-2025-period", "markdown": "https://wpnews.pro/news/pace-of-ai-capability-growth-is-in-oct-2026-is-2x-of-mar-2024-to-oct-2025-period.md", "text": "https://wpnews.pro/news/pace-of-ai-capability-growth-is-in-oct-2026-is-2x-of-mar-2024-to-oct-2025-period.txt", "jsonld": "https://wpnews.pro/news/pace-of-ai-capability-growth-is-in-oct-2026-is-2x-of-mar-2024-to-oct-2025-period.jsonld"}}