Analysis
METR says Anthropic redacted most of the evidence METR found most informative about AI speeding up Anthropic's own research, which METR puts at about 2x
METR's November 5 report: Anthropic improves models about 2x faster, mostly thanks to AI, but Anthropic redacted a majority of METR's most informative evidence.
- New AI technologies
- Models
- #METR
- #Anthropic
- #AI R&D acceleration
- #redaction
- #independent evaluation
- #Responsible Scaling Policy
- #Claude
On November 5, 2026, the evaluation group METR published its assessment of AI R&D acceleration at Anthropic. It concludes that Anthropic is likely improving its models on well-specified tasks at about twice the rate of March 2024 to October 2025, with most of the speed-up coming from AI use. It also says Anthropic redacted a majority of the evidence METR found most informative. Disclosure: Anthropic makes Claude, which helps draft articles on this site. We read METR's post only through search excerpts and press relays.
The essentials
- What "redacted" covers: a majority of the evidence METR found most informative, not a majority of all evidence. Reported cut items include per-stage acceleration estimates, bottleneck analyses, forecasts for future models and benchmark names.
- Who decided: Anthropic held redaction rights under its agreement with METR. Anthropic's stated reason is that the material is commercially sensitive and could help other AI developers.
- The conclusion is METR's judgment: about 2x on well-specified tasks, with high uncertainty (1.3x to 3x plausible). METR guesses a majority, perhaps about 80%, of the acceleration is due to AI.
- Next: Anthropic is said to plan its own report in the coming weeks.
Was the evidence censored?
The word fits only loosely. According to the coverage we read, Anthropic had unilateral redaction rights and used them heavily, but it did not control METR's conclusions. METR says the public version does not meet its own transparency standard, so readers must lean on METR's judgment more than usual. Anthropic co-founder Sam McCandlish is reported to have acknowledged on X that the cuts make it harder for others to check the work, while defending them as commercially sensitive.
What does the 2x figure mean?
METR estimates capability gains on well-specified tasks at about 2x the March 2024 to October 2025 pace. A trend fit with a breakpoint in October 2025 gives 1.9x (95% interval 1.5x to 2.8x), per press relays. Staff self-reports of AI speedup are far higher, up to 12x, but METR calls them loose evidence and notes that individual speedup is not whole-lab acceleration. METR also calls Anthropic's argument that its 2x threshold was not crossed overstated; Anthropic's August Risk Report said acceleration was below 2x through July 15.
What is not established
- What the redacted evidence shows: outsiders cannot check the per-stage estimates, bottleneck analysis or forecasts.
- Whether Anthropic's own upcoming report will disclose more, or reach the same figure.
- Whether acceleration will keep compounding: METR saw no major bottleneck over 12 months but says it is not confident and its view changed several times.
- Which independent reproduction, if any, is possible: none has been reported.
Summary table
| Point | What is stated as of November 6, 2026 | Source |
|---|---|---|
| Headline estimate | About 2x the March 2024 to October 2025 pace on well-specified tasks; 1.3x to 3x plausible | METR via Tech Times, DEV Community |
| Share due to AI | A majority, perhaps about 80% (METR's guess) | FourWeekMBA, METR excerpt |
| Redaction | A majority of the evidence METR found most informative | AlphaSignal, DEV Community |
| Reason given | Commercially sensitive; could help other developers | FourWeekMBA, AlphaSignal |
| Staff self-reports | Up to 12x, called loose evidence by METR | FourWeekMBA |
| Anthropic's follow-up | Own report planned in the coming weeks | FourWeekMBA |
FAQ
Did METR say AI is training Anthropic's models by itself?
No. Coverage says METR sees AI-assisted work, with only early signs of fully autonomous contributions.
Is the 2x figure confirmed by Anthropic?
Not in the sources we read. Anthropic's August Risk Report said below 2x through July 15; METR calls that argument overstated.
Can I check the numbers?
Only partly. The headline figures are public, but much of the evidence behind them is redacted.
Method and limits
Sources checked on November 6, 2026. We saw METR's post only through search excerpts, not in full, and other figures through press relays, which differ slightly in wording. Anthropic's own report and the unredacted METR report were not read. This is not investment or security advice.
Sources
- METR: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 (verified 2026-11-06)
- Tech Times: Anthropic AI Acceleration Hits 1.9x on Benchmarks; METR Calls Anthropic Threshold Defense Overstated (verified 2026-11-06)
- DEV Community: METR: Anthropic's Pace Roughly Doubled, Mostly From AI. Most Key Evidence Is Redacted (verified 2026-11-06)
- FourWeekMBA: METR: Anthropic Likely Improving Models at 2x Its Old Rate (verified 2026-11-06)
- Resultsense: METR finds AI is roughly doubling Anthropic's R&D pace (verified 2026-11-06)
- AlphaSignal: METR Finds Anthropic's AI-Assisted R&D Moving Twice as Fast (verified 2026-11-06)
Spotted an error in this article? Report it · Our corrections policy
