TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
METR’s independent assessment, published November 5, estimates that Anthropic’s progress on well-specified AI capability tasks accelerated about 1.9 times after October 2025. METR believes AI use likely accounts for much of the acceleration, but Anthropic redacted most of the evidence the evaluators considered most informative.
METR, a nonprofit that evaluates frontier AI systems, estimated on November 5 that Anthropic’s capability progress accelerated about 1.9 times after October 2025, based on an 11-week assessment of internal company data. The evaluators said AI likely drove much of the change, but Anthropic redacted a majority of the evidence METR found most informative, limiting what outsiders can independently examine.
METR compared Anthropic’s progress on a composite capability measure, called AECI, across two periods. It estimated a pace of about 15 points per year before October 2025 and about 29 points per year afterward. METR placed the estimated change point at October 7, 2025, and described the result as roughly two years of earlier progress occurring in one year. Its estimate of the acceleration was 1.9x, with a 95% confidence interval of 1.5x to 2.8x; METR said it would not be surprised by a value between 1.3x and 3x.
The result concerns measured capability progress on specified tasks. It does not mean every part of AI research or model development is moving twice as fast, nor does it establish that systems are autonomously designing their own successors. METR estimated that perhaps about 80% of the acceleration came from AI assistance and said it was confident the change did not result from faster growth in training compute. Those are the evaluators’ judgments, not a directly measured allocation of cause.
METR also reported that its estimate exceeds the 1.7x figure in Anthropic’s Haiku 5.5 system card, mainly because METR included newer, unreleased models. Looking across frontier models from all developers, rather than Anthropic alone, METR estimated a smaller change: from about 17 to 24 AECI points per year, or 1.4x. The assessment used benchmark results for unreleased models, training metrics, surveys and about 15 interviews with staff. METR said roughly 2.5 full-time-equivalent staff worked on it using Anthropic devices, and that it took no payment.
How Much Is AI Speeding Up? METR’s Anthropic Study
METR estimates Anthropic’s measured capability progress accelerated after October 2025. Redactions limit how much of the evidence outsiders can examine.
What METR measured
The estimate compares progress on specified capability tasks before and after an estimated October 7, 2025 change point.
- About 15 AECI points per year
- Earlier baseline pace
- About 29 AECI points per year
- Roughly two years of earlier progress in one
How the assessment worked
METR reviewed internal research and development activity over 11 weeks.
METR examined benchmark scores and training metrics.
The assessment included surveys and about 15 interviews.
The team estimated capability progress before and after the change point.
What the estimate does and does not show
Specified tasks
The finding concerns progress on METR’s capability measure and selected periods.
Autonomous successors
It does not show that AI systems are autonomously designing their own successors.
AI likely contributed
METR estimated AI assistance accounted for perhaps 80% of the acceleration; this was an evaluator judgment, not a measured allocation.
Evidence and policy questions
The headline estimate has implications for safety oversight, but important details remain open.
Why AI-Assisted Research Matters
If AI tools help researchers and engineers build more capable AI systems, improvements could compound: better models may assist with work on their successors. METR’s estimate gives policymakers, AI developers and safety researchers a measured indication of how quickly that feedback could be developing at one major lab. The report does not establish that fully autonomous self-improvement has arrived; Anthropic has said humans still set goals and that recursive self-improvement is not inevitable.
The estimate also bears on Anthropic’s own safety rules. Its Responsible Scaling Policy identifies a doubling in the pace of progress plausibly attributable to automated research or engineering as an early-warning threshold. METR said it was plausible that this threshold had been crossed by July 15, 2026. Anthropic separately wrote in a November 2 SEC filing that it “may imminently meet” its definition of dramatically accelerated progress. These statements refer to company policy and risk assessment, not a finding that a specific safety failure has occurred.
How much confidence readers should place in the headline estimate depends partly on evidence that is not public. METR’s access to internal data offers a different view from tests of released models, but redactions make it harder for outside researchers to reproduce the analysis or assess how the company’s data and choices shaped it.
Products Worth Considering
LG UltraGear 34GX90SA-W 34-inch WQHD Curved OLED Gaming Monitor
Bose SoundLink Flex Bluetooth Speaker (2nd Gen), Black
Kindle Paperwhite Signature Edition 32GB (2024) – Auto-adjusting front light, wireless charging, weeks of battery life – Metallic Black
As an affiliate, we earn on qualifying purchases.
How METR Assessed Anthropic
METR’s assessment is an embedded third-party review of internal research and development activity, rather than a test of a publicly released model alone. The team worked with Anthropic data from August 12 through October 29, 2026. METR says it examined benchmark scores, training metrics, staff surveys and interviews. Anthropic supplied laptops and free model tokens and retained the ability to redact non-public material.
The acceleration measure compares a recent rate with an earlier baseline. In METR’s framing, a 2x rate means that one year produces as much measured progress as two years did at the baseline pace. The estimate is specific to the capability index and periods studied; it is not a general measure of researcher productivity or a forecast of when AI will become fully self-improving.
Anthropic’s Responsible Scaling Policy version 3.4 took effect July 8, 2026. It names two automated research and development triggers: a system that could replace the company’s research staff at competitive cost, or a doubling in progress plausibly attributable in substantial part to automated research or engineering. The policy describes the threshold as a potential early warning.
““A 2x acceleration means 2 years of historical progress is made in 1 year.””
— METR
What the Redactions Leave Unknown
METR said Anthropic redacted a majority of the evidence it considered most informative. The public account therefore does not provide all the underlying material needed to independently evaluate the estimate, the quality of the measurements or the evaluators’ reasoning. METR also dropped a planned check of how Anthropic divides compute between safety work and capability development.
The report’s estimate is tied to Anthropic’s AECI measure, its selected models and an estimated October 2025 change point. METR’s lower 1.4x estimate across frontier developers suggests the result differs by scope. The available source material does not establish how much the redacted evidence could change the findings, whether other labs are experiencing comparable acceleration, or how much of the estimated AI contribution comes from particular tools and tasks.
It also remains unclear whether Anthropic has formally determined that its Responsible Scaling Policy threshold was met. METR said crossing it by July 15 was plausible; Anthropic’s filing said the company may imminently meet its definition. Neither statement, as reported here, confirms a formal determination.
Policy Review and More Evidence
The next key development is whether Anthropic formally assesses its automated research and development thresholds under its Responsible Scaling Policy and describes any resulting safeguards. The November 2 filing signals that the company sees the possibility of meeting its definition soon, while METR’s July estimate makes the timing of any internal review a point to watch.
Further independent evaluation will depend on what evidence Anthropic makes available and whether other researchers can inspect enough data to test METR’s methods. Public updates on model capability measures, AI use in research, and the division of resources between safety and capabilities could clarify whether the estimated pace persists. For now, the report provides a detailed but partly obscured snapshot of one company’s progress.
Key Questions
What did METR estimate?
METR estimated that Anthropic’s measured capability progress increased from about 15 to 29 AECI points per year after an estimated October 7, 2025 change point. It summarized that as a 1.9x acceleration, with a 95% confidence interval of 1.5x to 2.8x.
Does the report show that AI is building its own successor?
No. The report concerns AI assistance in research and engineering. It does not establish fully autonomous recursive self-improvement, in which an AI system independently designs and develops its successor.
How much of the acceleration did METR attribute to AI?
METR estimated that perhaps about 80% came from AI. That is the evaluators’ estimate of the cause, not a directly observed measure, and most of the evidence they found most informative was redacted.
Did METR find the same acceleration across the AI industry?
No. When METR compared frontier models from all developers, it estimated progress rose from about 17 to 24 AECI points per year, or 1.4x. The 1.9x estimate applies to Anthropic’s models in the assessment.
Has Anthropic confirmed it crossed its policy threshold?
The source material does not report a formal confirmation. METR said it was plausible the threshold had been crossed by July 15, while Anthropic wrote on November 2 that it may imminently meet its definition of dramatically accelerated progress.
Source: Thorsten Meyer AI
