METR published an independent assessment on November 5, 2026 concluding that, as of October 2026, Anthropic is likely improving model capabilities on well-specified tasks at about twice the rate seen between March 2024 and October 2025. Using Anthropic’s AECI (an Epoch Capabilities Index variant), METR estimates benchmark growth at 1.9× the pre-October 2025 rate (95% CI 1.5–2.8×). The team attributes a majority of the current acceleration to AI-assisted R&D, rules out faster training-compute growth as the cause, and warns that missing bottlenecks raise the risk of further speedups — while stressing high uncertainty and heavy Anthropic redactions in the public report.
Quick Take
- What: METR’s high-access review (Aug 12–Oct 29, 2026) finds ~2× capability progress pace at Anthropic on well-specified tasks vs the prior historical window.
- Headline metric: AECI trend acceleration 1.9× since October 2025 (CI 1.5–2.8×) — roughly two years of historical progress in one year.
- Attribution: METR guesses a majority (perhaps ≈80%) of current acceleration is from AI use; late-2025 speedup had larger non-AI factors; not from training-compute growth.
- Why it matters: An outside evaluator with internal access is saying recursive AI R&D acceleration is no longer only a theory slide.
What happened
METR (@METR_Evals) announced the public report summarizing months of work inside Anthropic. The full write-up lives at metr.org. Anthropic co-founder Sam McCandlish replied that METR had sensitive internal data and independence, that many conclusions match his own, and that Anthropic redacted heavily because sharing “exactly what is driving the speedup” could hand rivals a roadmap. AP, The Verge and others carried the ~2× framing with METR’s caveats attached.
Background: why this assessment exists
Anthropic lists AI R&D acceleration among key risk indicators in its Responsible Scaling Policy. Outside observers usually see only system cards and public benches. METR’s engagement was scoped to whether AI is accelerating AI development, with access to internal benchmarks, training-domain compute metrics, staff interviews and related analyses — short of full employee-like roaming access or sampling non-public models. A compute-allocation verification piece was canceled mid-engagement; METR discloses that change.
How METR measured acceleration
METR fits trends on AECI scores for public and internal models (including unreleased “Model 2–4” snapshots dated by internal availability). A learned breakpoint around early October 2025 separates slower and faster regimes. Robustness checks across link functions, snapshots and benchmark subsets mostly keep acceleration near ~2× (with some specs as low as 1.1× or as high as ~2.9×). Interviews and Slack-poll uplift self-reports are treated as supportive but loose evidence. Fully autonomous contributions to the most important training steps are not yet dominant, METR says, though “signs of life” appear.
| Finding | METR public report |
|---|---|
| Pace (subjective) | ~2× vs Mar 2024–Oct 2025 on well-specified tasks; uncertainty band ~1.3–3× |
| AECI acceleration | 1.9× since Oct 2025 (95% CI 1.5–2.8×) |
| Compute cause? | Ruled out: not from faster training-compute growth |
| AI attribution (current) | Majority; METR rough guess ≈80% |
| Bottlenecks (next 12 mo) | No clear blockers found; low confidence |
| Transparency | Majority of most informative evidence redacted by Anthropic |
Pros and cons
For users
- Pro: Faster Claude-class progress may show up as better products sooner if the AECI slope maps to real tasks.
- Con: Acceleration on clean, well-specified tasks may overstate messy real-world reliability.
For developers
- Pro: Confirms that AI coding/experiment assistance inside a frontier lab is already a first-order R&D input.
- Con: Public builders cannot see the redacted metrics that most informed METR’s risk view.
For companies
- Pro: A template for demanding independent R&D-acceleration assessments from frontier vendors.
- Con: Heavy redactions mean boards and regulators still lean on evaluator judgment rather than inspectable evidence.
What X is saying
The thread mixed safety alarm with respect for the access METR received.
Since August, a team within METR has worked with internal Anthropic information to assess whether AI itself is accelerating AI development. We are now publishing a report summarizing the results of this risk assessment.
— METR (@METR_Evals) November 5, 2026 · View on X
We redacted a lot, and I accept that this makes it harder for others to check METR’s work. But we determined that sharing exactly what is driving the speedup isn’t the right thing to do…
— Sam McCandlish (@Sam_McCandlish) November 5, 2026 · View on X
METR’s report has a ton of details on Anthropic’s internal models, including “Model 4”, which finished training in September-October…
— prinz (@deredleritt3r) November 5, 2026 · View on X
Anthropic made roughly two years of progress in the last year. METR… estimates benchmark scores are growing at 1.9x the pre-Oct-2025 rate… confident the acceleration isn’t from more compute.
— Jay Zhou (@jayzhoupro) November 6, 2026 · View on X
What Grok says
Grok summarized the assessment’s core numbers — ~2× pace, AECI 1.9× since Oct 2025, CI 1.5–2.8×, majority AI attribution, not compute scaling, no clear near-term bottlenecks — matching the public report’s executive points.
Critical analysis
METR is careful in ways the viral summary is not. The ~2× claim is subjective-consensus language with a 1.3–3× “would not be surprised” band. AECI could overfit verifiable tasks; Anthropic’s earlier subfrontier slope could inflate apparent acceleration; non-AI factors mattered in late 2025; and the most informative evidence is redacted. Still, ruling out training-compute acceleration while seeing sustained AECI steepening is a big deal. McCandlish’s defense of redactions is coherent as a race-dynamics argument and costly as a transparency argument — both can be true. The public should treat this as strong evidence that AI-assisted R&D is already bending a frontier lab’s capability curve, not as proof of an imminent fully autonomous intelligence explosion.
What’s still open
- Will Anthropic’s forthcoming internal progress report reconcile with METR’s 1.9×?
- Can future assessments publish more evidence without a rival roadmap?
- Does acceleration transfer to messy, long-horizon tasks beyond AECI?
- What happens when autonomous experiment loops move from “signs of life” to default?
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis is based on METR’s November 5, 2026 public report, METR’s and Sam McCandlish’s posts on X, and secondary wire coverage. Redacted sections are discussed only as METR describes them. More on Anthropic.
i10x Perspective
Recursive self-improvement stopped being a sci-fi slide when an independent evaluator with internal access said a frontier lab’s well-specified capabilities are compounding near 2× historical speed — mostly via AI helping humans research AI. The next transparency fight is whether the public gets enough unredacted evidence to audit that claim before the slope steepens again.
