Anthropic is improving its AI models on measurable tasks at nearly twice its previous pace, according to an independent investigation by Model Evaluation and Threat Research (METR). The researchers believe AI is responsible for much of the acceleration, but their findings do not show that Claude can autonomously develop and train its successors.
METR published its assessment on 5 November 2026, after examining Anthropic’s internal research data and interviewing employees between 12 August and 29 October. The findings received renewed attention in The Verge on 7 November. We examined the underlying methodology, the limits on what METR could disclose, and what the results mean for developers, AI buyers and safety teams.
What the 1.9x acceleration figure actually measures
METR measured model capability trends using Anthropic’s version of the Epoch Capabilities Index (AECI), a composite of benchmarks covering tasks with reasonably clear success criteria. Since October 2025, those scores have grown at an estimated 1.9 times the earlier rate, with a reported 95% confidence interval of 1.5x to 2.8x.
In plain terms, that suggests Anthropic made roughly two years of historical benchmark progress in the most recent year. It does not mean Claude has become 90% more intelligent, completes every task 1.9 times faster, or delivers twice the commercial value. METR’s alternative modelling choices produced acceleration estimates ranging from 1.1x to 2.9x.
| Finding | Reported figure | How to interpret it |
|---|---|---|
| Benchmark progress | 1.9x faster growth | An estimated change in the rate of improvement on selected tasks, not an absolute capability score. |
| Share attributed to AI | Perhaps 80% | METR’s uncertain judgment about the current acceleration, not a controlled causal measurement. |
| Researcher productivity | About 12x self-reported uplift | Informal employee polling for a strong internal model, not verified end-to-end research throughput. |
| Fully autonomous core research | Not yet established | Human researchers remain essential to the most important parts of training. |
The distinction between these numbers is essential. A researcher feeling twelve times more productive does not imply that a laboratory’s models improve twelve times faster. Output depends on experiment quality, training resources, coordination and how much useful progress survives review.
Claude is shortening the route from research idea to experiment
The most concrete evidence concerns AI-assisted research workflows. Interviews describe Claude writing and modifying code, implementing experimental ideas and making it easier to test more approaches in parallel. One researcher reported that coding was no longer the main constraint on their work; generating worthwhile ideas and finding compute had become more limiting.
That change can affect a research programme even when a person still makes every consequential decision. If engineers can implement a speculative experiment quickly, they can afford to explore ideas that previously required days of setup. Unsuccessful experiments may become cheaper to try, while human researchers still decide which results are credible and worth pursuing.
There is a catch. More experiments are not automatically better experiments. A team can increase throughput while also creating more results to validate, more infrastructure to monitor and more opportunities to optimise towards its own benchmarks. METR did not publish a reproducible, stage-by-stage account of how much AI contributed to pretraining, reinforcement learning or evaluation.
The investigators also say faster growth in training compute does not explain the observed acceleration. That finding rules out one particular explanation, not the importance of compute itself or other contributors such as hiring, improved data and algorithmic research.
The report’s most significant limitation is what Anthropic withheld
METR received unusually detailed access to internal benchmarks, unpublished model results, research inputs, and employee interviews. Yet this was not unrestricted access. Anthropic mediated information requests, the evaluators did not directly sample unreleased models, and the company retained the right to redact non-public material from the published assessment.
According to METR’s original assessment of Anthropic’s AI research acceleration, Anthropic removed most of the evidence that had most influenced the researchers’ overall risk judgement. The public cannot independently reconstruct the detailed attribution of gains to AI, audit the proposed bottlenecks or examine the full forecasts for future models.
A further complication is that Claude was involved in substantial parts of the information gathering and analysis, including METR’s own work. The evaluators say they found no evidence of systematic distortion when cross-checking against human interviews. Nevertheless, an independent assessment that relies substantially on the system being assessed needs especially clear provenance, reproducible calculations, and human verification.
Outside reactions have repeatedly focused on this transparency problem: independent investigators may reach their own conclusions, but readers cannot independently check conclusions supported by withheld evidence. That is a legitimate limitation, not a reason to dismiss all the findings.
Anthropic acknowledges the speedup but defends the redactions
Anthropic co-founder Sam McCandlish responded on 5 November, saying AI was helping accelerate the company’s research and that many of METR’s conclusions matched his understanding. He also acknowledged the difficulty of verifying a heavily redacted report, arguing that disclosing the precise drivers of faster development could expose commercial secrets and help competitors accelerate their own research.
In a separate review of Anthropic’s August risk reporting, METR questioned whether the company had presented the possibility of crossing its AI-driven R&D acceleration warning threshold clearly enough. The roughly 2x threshold is an early-warning indicator under Anthropic’s Responsible Scaling Policy, not proof that catastrophic risk has materialised. METR did not conclude that Anthropic’s July models could replace its full research workforce.
The policy question is therefore specific: can an independent reviewer verify when rapid AI-assisted progress should trigger stronger controls, even where the company disputes the causal attribution or withholds the relevant metrics?
What should developers and AI governance teams do differently?
For companies buying or integrating AI, the immediate lesson is not to forecast a new model every few weeks. Instead, stop treating a vendor’s current benchmark ranking as a stable procurement decision. A faster research cycle can shorten the useful life of a model comparison while increasing the chance that organisations deploy updates they have not properly tested.
- Test each meaningful model release on held-out work. Include tasks with ambiguous requirements, long-running tool use and failure recovery, rather than relying only on published benchmark scores. DIY AI’s AI model comparison provides a separate starting point for evaluating cost against successful task completion.
- Record cost and review overhead together. For coding agents, log retries, tokens, engineer approval time and regressions. Subscription capacity and API metering create different economics, as detailed in our Claude Code pricing guide.
- Keep the research and deployment loop auditable. Version-pin models where possible, preserve experiment inputs and outputs, isolate agent permissions and retain a human approval gate before changes reach production.
- Ask what an external audit could actually reproduce. A vendor’s willingness to invite evaluators is positive, but disclose whether reviewers saw raw logs, controlled the tests, examined failures and can publish the evidence that shaped their conclusions.
What would prove AI is improving itself without researchers?
METR’s findings support a narrower conclusion than the strongest self-improvement headlines: AI appears to be accelerating human-led AI research, and further acceleration is plausible. The report does not establish a closed loop in which an AI system independently designs a stronger successor, trains it, evaluates its limitations and repeats the process without significant human direction.
The next evidence worth watching is independent verification of autonomous contributions to core training decisions, improvements on new tasks outside the benchmarks used to guide development, and clear measurements of how much faster research becomes compared with a credible no-AI baseline. METR found no convincing near-term bottleneck, but explicitly warned that important constraints may have been missed.
For now, the defensible headline is that Anthropic’s measured model progress has accelerated sharply and AI assistance is probably an important cause. The unresolved issue is whether the public and independent safety assessors can inspect enough evidence to recognise when useful automation becomes a self-reinforcing capability race.
