Accelerating AI development at Anthropic: what the METR estimate shows

The METR estimate describes not “Anthropic’s intelligence” as a single property, but a change in the rate at which the capabilities of the company’s models are growing. The report summary claims that this rate increased noticeably after October 2025, but the available materials do not make it possible to independently verify all the original data and hidden appendices of the study.

Brief conclusion: an acceleration figure of around 1,9 times appears significant, but does not prove the emergence of superintelligence, complete AI self-improvement, or an inevitable transition to autonomous research.

What the Anthropic IC index measures

The Anthropic IC index, as described in the report summary, combines the results of many tests into a single metric. METR used it to assess how quickly the capabilities of Anthropic models were changing across different areas, rather than to measure general intelligence directly.

Such an index is useful for analyzing trends: a researcher compares values at different points in time and estimates the slope of the curve. The final figure depends on the composition of the tests, their difficulty, the thresholds for successful completion, and how the methodology accounts for models that are already close to the upper limit of a particular task.

IC cannot be directly interpreted as a universal intelligence scale. A model may improve its performance on tasks with automated grading while changing very little in open-ended scientific work, where formulating the problem, choosing a hypothesis, and evaluating unexpected results are required.

How Anthropic’s development rate changed

Before October 2025, Anthropic models, according to the METR summary, were gaining about 15 IC points per year. After the presumed turning point, the estimate rose to approximately 29 points per year.

Mathematical analysis dated the turning point to 7 October 2025. A specific model with a new level of capability did not necessarily launch on that day: the statistical procedure looks for a point where the trend changed and associates it with a period, not necessarily with a single release or technological decision.

The summary also gives another comparison: in the year after the turning point, Anthropic achieved approximately the same capability gain that had previously required two years. The comparison refers to the period from October 2023 to October 2025, so it cannot be turned into a forecast that every following twelve-month period will bring another two former annual steps for the company.

Why the estimate is not exactly 1,9

The number 1,9 is the central estimate. For the acceleration, METR reported an interval from 1,5 to 2,8 times with 95 percent statistical confidence. It shows that uncertainty around the result is substantial, not that every value within the range is equally likely.

The result is affected by the choice of statistical model and the composition of the tests. The summary states that the logistic model yielded an acceleration of 1,9 times, the probit model yielded about 1,8, other methods produced higher values, and under certain assumptions the figure fell to 1,1 times. The researchers retained the standard logistic approach because the different models appeared to describe the validation data similarly, but acknowledged the uncertainty of the result.

Robustness checks do not eliminate methodological risk. After excluding individual tests, including OSWorld and MMU, or tasks on which models had already exceeded the 85 percent threshold, the estimate remained approximately in the range of 1,9–2,2 times. The calculation still does not become direct observation of the pace of research inside the company.

Why Anthropic and the industry produce different figures

In technical documentation for Haiku 5.5, Anthropic had previously estimated the acceleration at approximately 1,7 times. METR obtained about 1,9 because it included unpublished models and took into account when they became available to most employees, rather than only the date of public release.

The difference between an estimate for one company and an estimate for the entire industry is even more pronounced. When leading models from OpenAI, Google, xAI, and other developers were included in the calculation, the growth rate increased from 17 to 24 IC points per year. This corresponds to acceleration of approximately 1,4 times, which is lower than the Anthropic result.

There are several possible explanations for this discrepancy, and according to the summary, METR did not choose one of them. Anthropic may have been catching up with competitors more quickly, other companies may have optimized their models for particular tests more effectively at an earlier stage, or the differences may have resulted from the composition of the data and the availability of internal versions.

Does AI help create the next models?

The report suggests the interpretation that AI is already accelerating the work of specialists developing the next generation of models. In interviews, employees described a shift from small code fragments under constant supervision to running several agents in parallel and assigning them different tasks.

According to employees’ accounts, AI handles code writing and debugging, infrastructure preparation, and part of the experimental work. One participant estimated that their own productivity had more than doubled; another attributed a significant portion of the benefit to reducing the time between an idea and an experiment; and separate estimates of AI’s contribution ranged from 15–20 to 50–80 percent.

These accounts should not be conflated with measuring the rate of model development. An employee’s subjective assessment reflects their perception of personal effectiveness, but does not separate AI’s contribution from hiring new specialists, increasing computing resources, process changes, and choosing tasks successfully.

According to the summary, METR does not consider fully autonomous research to have been achieved. AI is still unable, without human involvement, to make a truly significant contribution to all the most important stages of training leading models, including formulating the research task and checking that the discovered solution works beyond a convenient test.

What limits the reliability of the conclusion

The public version of the report does not contain some of the most informative materials. Anthropic was given the right to remove confidential information, and detailed estimates of algorithmic progress, possible limitations, and forecasts of future capabilities were removed from the text. The reader sees the methodology and final figures without the level of detail needed to fully reproduce the analysis.

There is also a chronological problem. The provided materials simultaneously mention estimates from before October 2025, data validation in 2026, and models with numbers assigned dates in 2026. Without an additional original document, these details cannot be combined into one consistent timeline, so the names and characteristics of the secret models should be considered unconfirmed.

• METR did not have direct access to the alleged Model 4 and could not test it in dialogue or on tasks.
• Reports about the scores of Model 4, its timeline, and its possible connection to other names remain rumors.
• Predictions that superintelligence will appear within several months or one or two years do not follow from the acceleration estimate.
• The estimate of an AI contribution of about 80 percent is described as preliminary and has not been confirmed by comparison with an otherwise similar company without AI.
• According to the available summary, the report itself does not fully meet the transparency requirements for independent evaluations.

Which conclusions are supported and which are interpretations

What the METR estimate means for AI-team leaders

The practical significance of the result is not the 1,9 multiplier, but the need to measure which parts of the development cycle are actually accelerating. It is important for a team to distinguish the speed of writing code, the number of experiments conducted, the quality of hypotheses, the cost of computing, the time required for verification, and the number of errors discovered only after deployment.

If AI accelerates the generation of experiments, the bottleneck becomes not programming but selecting good ideas, access to computing, and independent verification of results. At the same time, the risk increases that tests will fall behind the models’ actual capabilities: tasks that recently distinguished strong systems from weak ones may become too easy and cease to show further progress.

For researchers, this means the need to repeat measurements on new task sets, disclose how sensitive the result is to the statistical model, and separately assess the contributions of AI, workforce growth, and computing resources. Safety leaders need a different answer: are evaluations, red teaming, and analysis of new capabilities keeping pace with the rate at which models begin helping to create their successors?

FAQ

Did Anthropic really start developing AI 1,9 times faster?

According to the estimate attributed to METR, the growth rate of the IC metric increased by approximately 1,9 times. This is not a measure of the company’s overall operating speed and does not guarantee the same acceleration in the future.

Does the report prove AI self-improvement?

The report supports a more cautious conclusion: AI may have accelerated certain stages of the work done by people creating new models. Evidence of fully autonomous recursive self-improvement is absent from the available summary.

Why did the figure for the entire industry come out to 1,4 times?

After models from other developers were included, the rate changed from 17 to 24 IC points per year. The broader sample smoothed out the Anthropic result and showed that the acceleration cannot be extended to the entire market without additional justification.

How accurate is the estimate of 1,5 to 2,8 times?

This is an interval of statistical uncertainty, not a promised range of future growth. It shows that the central estimate of 1,9 depends on the data, tests, and selected analytical method.

Conclusion

The METR report summary provides a serious signal: the rate at which Anthropic models improve may have increased after October 2025, and AI is probably already affecting the work of research teams. The strength of the conclusion is limited by closed data, incomplete reproducibility, dependence on the selected index, and the blending of measured results with interviews and forecasts.

The correct formulation is as follows: METR estimated that the growth rate of Anthropic’s capabilities accelerated by approximately 1,9 times, while for the entire industry it obtained about 1,4 times. This is an important empirical result for further investigation, but not proof that superintelligence has emerged and not a precise timetable for the future development of AI.