# Is AI speeding up AI research? What METR found at Anthropic, and what it did not

By Spezia. Published November 10, 2026. Updated November 10, 2026. Data as of November 10, 2026.
Canonical page: https://spezia.ai/blog/metr-ai-speeding-up-ai-research

## The short answer

METR, an independent evaluator, estimated on November 5, 2026 that Anthropic's benchmark scores have grown at about 1.9 times the earlier rate since October 2025, with a 95% range of 1.5 to 2.8. That measures benchmark scores on well-specified tasks, not research output. METR guessed that perhaps 80% of the current acceleration is due to AI, with substantial uncertainty, and said Anthropic redacted from the public report most of the evidence METR found most informative.

- METR estimated that Anthropic's benchmark scores grew at about 1.9 times the earlier rate after October 2025, with a 95% range of 1.5 to 2.8 times.
- The 1.9 figure is a growth rate in benchmark scores on well-specified tasks. It is not a measure of research output or of how much faster people work.
- The "about 80%" is a guess METR labels as uncertain, and it applies to the October 2026 rate, not the jump in late 2025.
- Anthropic redacted most of the evidence METR found most informative, so readers must rely on METR's judgment.
- METR's time-horizon page, which it no longer updates, lists an early version of Claude Mythos Preview at least 16 hours, and METR says its test cannot reliably measure above that.
- METR's 2025 finding that developers were 19% slower with AI is now labeled out of date by METR itself.

## What did METR find?

METR, a nonprofit that evaluates AI systems, published a report on November 5, 2026 titled an independent assessment of AI research acceleration at Anthropic. From August 12 to October 29, METR staff ran a high-access evaluation at Anthropic, with about 2.5 full-time-equivalent staff in total.

The headline is a rate. Benchmark scores for Anthropic's models, measured on a composite index Anthropic calls AECI, have grown since October 2025 at an estimated **1.9 times** the earlier rate, with a 95% confidence interval of 1.5 to 2.8 times. METR describes that as roughly two years of historical progress in the last year. The earlier rate is the trend before October 2025, and the index includes unreleased internal models.

- **1.9x**: estimated growth in benchmark scores after October 2025, against the earlier rate
- **1.5 to 2.8x**: METR's 95% confidence interval
- **1.4x**: the same estimate using frontier models from all developers
- **1.1x to 2.9x**: the range across METR's robustness checks

The score trend rose from about 15 index points a year before October 2025 to about 29 a year after, according to METR's report. METR offers two readings of the gap between the 1.4 and 1.9 figures and does not choose between them.

## What does 1.9 times measure, and what does it not?

It measures how fast benchmark scores are rising. It is not a count of researcher hours or ideas, and it is not how much faster people work. METR's own overall judgment is that Anthropic is improving model capabilities on well-specified tasks at around twice the rate it did in 2024 and 2025. Well-specified tasks are ones with clear goals and checks.

Two numbers in the headlines need care.

- **"Research pace doubled."** METR frames its result as the pace of capabilities progress, measured by benchmark scores. The word doubled comes from the roughly two-times result.
- **The range.** METR's confidence interval covers only uncertainty in the scores, not which models were sampled. Across different statistical choices, the estimates run from 1.1 to 2.9 times. METR says it would not be surprised if the true figure were closer to 1.3 or to 3 times.

METR also reports that it received convincing evidence the acceleration was not due to faster growth in the scale of training compute.

## Where does the "80%" come from?

From a judgment, not a measurement. METR wrote that it guesses a majority, perhaps about 80%, of the acceleration seen in October 2026 is due to AI, and gave that guess with substantial uncertainty. In other words, the claim is that AI is now helping to build better AI.

The qualifier matters. METR also said the acceleration in late 2025 was likely influenced by factors other than AI. So the 80% applies to the October 2026 rate. Other views are in METR's interview notes: one person guessed that 15% to 20% is due to AI, and another said hiring was probably the biggest reason.

Staff also reported speedups in internal polls that averaged about twelve times for the most capable internal model. METR called that very loose evidence.

## What are the limits?

METR lists them itself.

- **Redactions.** Anthropic redacted most of the evidence METR found most informative. METR says this means the public has to rely on its judgment much more than would be ideal.
- **Saturating benchmarks.** The most capable internal model scores above 50% on all 15 benchmarks it was tested on, and METR says that makes it somewhat nervous about its ability to measure the index now and in future.
- **Other causes.** Non-AI factors may explain most of the change, it may apply only to well-specified tasks, and Anthropic's rate in 2023 to 2025 may have been unusually low.
- **No access to non-public models.** METR did not request or receive access to sample from them.
- **Conflicts.** No fees were paid either way. Anthropic provided laptops and free tokens, and one METR team member discussed a role with Anthropic and declined an offer.
- **Autonomy.** METR describes the AI contributions to the most important parts of model training as not yet highly significant and fully autonomous. Researchers described signs of life.

A companion review of the research section of Anthropic's August 2026 risk report found that the report said acceleration was less than a factor of two. METR's own calculation from the redacted figures came to 1.97, which it said implies around two. Anthropic's filing addresses the same policy question, as described in [Anthropic's IPO filing, explained](https://spezia.ai/blog/anthropic-s-1-explained).

## What did Anthropic say?

Sam McCandlish, an Anthropic co-founder, replied in a thread on X. He said many of METR's conclusions match his own, and that AI is speeding up Anthropic's research, but it is hard to say exactly how much. He accepted that the redactions make METR's work harder for others to check, and said sharing what drives the speedup is commercially sensitive. He said Anthropic plans to share its own analysis in the coming weeks, including where it stands against its Responsible Scaling Policy, and that METR had complete independence to reach its own conclusions.

## What is METR's time horizon, and where does it stand?

The time horizon is a different measure. METR defines it as the length of task, measured by how long a human expert takes, that an AI agent is predicted to complete successfully at a given reliability. At 50% reliability it is the task length the agent completes half the time. It measures task difficulty, not how long the AI works.

| Model | Date in METR's data | 50% time horizon, hours | 50% range, hours |
| --- | --- | ---: | ---: |
| Claude Mythos Preview (early version) | April 7, 2026 | At least 16 | 8.5 to 55.1 |
| Claude Opus 4.6 | February 5, 2026 | 12.0 | 5.3 to 60.6 |
| Gemini 3.1 Pro | February 19, 2026 | 6.4 | 3.9 to 11.6 |
| GPT-5.2 | December 11, 2025 | 5.9 | 3.3 to 13.6 |
| GPT-5.3 Codex | February 5, 2026 | 5.8 | 3.2 to 13.6 |
| GPT-5.4 | March 5, 2026 | 5.7 | 3.1 to 12.8 |
| Claude Opus 4.5 | November 24, 2025 | 4.9 | 2.7 to 10.4 |

METR says measurements above 16 hours are unreliable with its current task suite, because only 5 of its 228 tasks are estimated at 16 hours or more. For that reason it does not highlight exact values above 16 hours, and says the most capable agents essentially saturated its time horizon benchmark. METR's own data file puts the doubling time of the horizon at about 129 days from 2023 on, using only measurements at or below 16 hours, with a range of 104 to 158 days. Its 2025 headline figure was seven months. METR's page says it is no longer actively updated, with a last update on September 8, 2026.

## Does AI make developers faster?

METR's 2025 randomized trial found that 16 experienced open-source developers took 19% longer with AI tools on 246 real issues, though they had expected to be faster. METR now labels that result out of date. In a February 2026 update it said developers are likely more sped up now. It also said its new estimates are unreliable, because many developers declined to work without AI, which biases the comparison. Do not quote the 19% as a current figure.

## How should you read these numbers?

Five checks help.

1. **What was measured?** A benchmark score, a task length and a worker's output are three different things.
2. **What is the comparison?** A rate is only meaningful against an earlier rate, and here the baseline is the trend before October 2025.
3. **Is it a measurement or a guess?** The 80% is a stated guess.
4. **What was withheld?** Most of the informative evidence was redacted.
5. **Who gained access?** METR had high access but did not sample non-public models.

Open AI Progress for releases, benchmarks and an outlook for the next 3 to 24 months, with the source on every figure. Ask Spezia what a new result means for your plans. [Track how fast AI is improving](https://spezia.ai/tools/ai-progress)

## How this was made

Every figure is read from METR's own publications and data file, and from Anthropic's response as posted by its co-founder. Wording is checked against what METR says it measured. Where a figure in the news differs from METR's wording, the difference is explained. Spezia makes no forecast about AI progress.

## Frequently asked questions

### Is AI speeding up AI research?

METR estimates that Anthropic's benchmark scores have grown at about 1.9 times the earlier rate since October 2025, and guessed that perhaps 80% of the current acceleration is due to AI. That is a judgment with substantial uncertainty, and it measures benchmark scores, not research output.

### What does METR's 1.9x mean?

It is the estimated growth rate of benchmark scores for Anthropic's models after October 2025, compared with the earlier trend. The 95% range is 1.5 to 2.8 times. It is about well-specified tasks and is not a measure of how fast people work.

### What is METR's time horizon?

The length of task, measured by how long a human expert takes, that an AI agent can complete at a given reliability. The highest figure on METR's page, which is no longer updated, is at least 16 hours at 50% reliability, for an early version of Claude Mythos Preview. METR says its suite cannot reliably measure above 16 hours.

### Did Anthropic agree with METR's findings?

In part. Co-founder Sam McCandlish said many of METR's conclusions match his own and that AI is speeding up Anthropic's research, though it is hard to say by how much. He accepted that the redactions make METR's work harder to check.

### Why did Anthropic redact evidence?

Anthropic said sharing what drives the speedup is commercially sensitive and could give a roadmap to others. METR says the redactions removed most of the evidence it found most informative, so the public must rely on its judgment.

### Does AI make developers slower?

METR's 2025 trial found experienced developers took 19% longer with AI tools. METR now calls that out of date and says developers are likely more sped up, but says its newer estimates are unreliable because many declined to work without AI.

## Sources

1. [Independent assessment of AI R&D acceleration at Anthropic in October 2026](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/), METR
2. [Review of the risks from automated R&D section in Anthropic's Risk Report (August 2026)](https://metr.org/blog/2026-11-05-rd-section-anthropic-risk-report-august-2026-review/), METR
3. [Task-completion time horizons of frontier AI models](https://metr.org/time-horizons/), METR
4. [METR time horizon benchmark results, version 1.1 (data file)](https://metr.org/assets/benchmark_results_1_1.yaml), METR
5. [Frontier risk report, February to March 2026](https://metr.org/blog/2026-05-19-frontier-risk-report/), METR
6. [Early-2025 AI and experienced open-source developer productivity study](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/), METR
7. [Uplift update, February 2026](https://metr.org/blog/2026-02-24-uplift-update/), METR
8. [Sam McCandlish's thread responding to METR's report](https://x.com/Sam_McCandlish/status/2118422266669371519), X
9. [METR finds Anthropic may have crossed its own AI acceleration warning threshold back in July](https://the-decoder.com/metr-finds-anthropic-may-have-crossed-its-own-ai-acceleration-warning-threshold-back-in-july/), The Decoder

## Cite this

Spezia. "Is AI speeding up AI research? What METR found at Anthropic, and what it did not" Spezia, November 10, 2026. https://spezia.ai/blog/metr-ai-speeding-up-ai-research