Anthropic’s internal Model 4 appears in a METR assessment
A previously undisclosed Anthropic model known as Model 4 has appeared in a public assessment by the third-party evaluation group METR. The report says the model was trained between September and October 2025 and reached 178.6 on Anthropic’s internal Aggregate Capability Index, or AECI. That score is above the 174.5 recorded by the public Opus 5.5 model and the 172.5 recorded by Mythos 5.1.

METR said Anthropic’s rate of improvement on clearly defined tasks had reached twice the rate seen across the March 2024 to October 2025 period by October 2025. The report treats October 2025 as the point when the model-development curve entered a faster phase, rather than describing the change as a single jump.
The report’s appendix lists four undisclosed Anthropic models: Unreleased 2022, trained in 2022; Model 2, trained in July 2026; Model 3, trained in August and September; and Model 4, trained in September and October. The source material says some online users have speculated that Model 3 could be the rumored Mythos 5.5, but the report does not confirm that identification.
Model 4 cleared 50% on all 15 benchmarks
AECI converts results from multiple benchmarks into a combined capability score. Model 4 exceeded 50% accuracy on each of the 15 benchmarks listed in the assessment. Nine months earlier, Opus 4.6 had still fallen below that mark on several benchmarks.

METR also examined the test set and found that three of the 15 benchmarks contained problems linked to unusually difficult long-tail items or damaged questions. Removing those three benchmarks would add 1.1 points to Model 4’s AECI result, lifting it to 179.7.
The evaluation may not have fully measured Model 4’s capacity. A standard benchmark run used an average of about 2.5 million input tokens and generated more than 70,000 output tokens for a single task. A coding task handled by an agent used 10 million input tokens and produced about 94,000 output tokens.
Annual AECI gains rose from 15 points to 29
METR identifies Oct. 7, 2025, as a historical dividing point in Anthropic’s capability curve. Before that date, Anthropic models gained roughly 15 AECI points per year. From October 2025 onward, the annual increase rose to 29 points, or about 1.9 times the earlier rate. The report gives a 95% confidence interval of 1.5 times to 2.8 times.

In a risk report published in August 2025, Anthropic presented a milder explanation: the change might have been a one-time step caused by new tools and new evaluation methods. METR tested that explanation against a model of sustained acceleration.
Researchers used data from Mythos 5 and earlier models to draw two forecast lines. They then ran a blind comparison against Model 2, Model 3, Opus 5.5 and Model 4. In 81% of the statistical samples, the sustained-acceleration forecast had an error of 1.9 points, compared with 3.9 points for the one-time-jump forecast.
The report uses an exponential-growth example to illustrate the effect of a faster curve. A metric that originally tripled each year could, under a twofold acceleration assumption, increase ninefold per year, or 3 squared. That is a mathematical example in the report, not a measurement of a specific model metric.

AI assistance accounted for about 80% of the acceleration
METR separated training compute, algorithms and staffing when examining the source of the faster development pace. The group ruled out a sharp increase in training compute as the main explanation, saying the growth rate of training compute after October 2025 was actually slower than before.
METR estimated that about 80% of the roughly twofold acceleration in research came from AI assistance. The remaining portion was attributed to a larger researcher headcount and expanded inference compute. The report’s description is direct: stronger Claude systems are helping researchers build stronger AI systems.
Internal employee surveys point in the same direction. In spring 2025, employees using Claude 3.7 Sonnet and Opus 4 reported productivity gains of about 1.3 times to 1.4 times. With Mythos 5 in June 2026, the self-reported gain rose to around five times. By October 2026, employees using Model 4 reported research acceleration of about 12 times. An exponential fit corresponds to a 3.9-fold annual increase.

Researchers began directing multiple agents at once
To examine the reported 12-fold gain, METR interviewed 11 Anthropic technical researchers in late September 2025. Researcher J described the workflow in three stages.
- At first, researchers used Claude mainly as an autocomplete tool for small, one-off scripts.
- In the second stage, a researcher supervised one Claude instance as it developed a complete feature.
- In the current stage, researchers direct groups of agents and run experiments in parallel.
Researchers C, D and E said Anthropic launched an internal incentive program in 2026 to encourage employees to hand previously impractical ideas to AI agents. Researchers could submit 10 research ideas to the internal agent system at the same time and see which ones produced results.
Researcher A said coding experiments by hand had been time-consuming. With Claude’s help, the time from an initial idea to a working experiment fell by 80%. The researcher said the remaining limits were ideas and compute, not code writing. Researcher B estimated a threefold overall productivity gain and compared it to suddenly having several on-call postdoctoral researchers and graduate students.

METR said AI has not yet taken full independent control of the design of the most central model-training work. Several researchers interviewed by METR nevertheless confirmed that AI had begun making visible autonomous contributions to core training.
Anthropic’s public version removed part of the evidence
In an official post, METR said it had submitted a complete report of roughly 100 pages to Anthropic. The public version removed much of the material described as the most information-rich evidence. Numerous sections carry redacted labels.
A table that had been used to forecast capability scores for Anthropic’s new models over the following several months, along with the related analysis, was removed as a block. METR said it saw no clear bottleneck in algorithms, compute or data that would stop the acceleration machine within the next 12 months.

The source material also says Anthropic updated its S-1 filing before an initial public offering and added a risk disclosure. An internal estimate cited in the article says future models could soon cross a threshold at which AI significantly accelerates research and development, completing in one year work that previously took two.
METR researcher Ryan Greenblatt described the report as a major signal for achieving artificial general intelligence or ASI within one to two years. The discussion was also linked to the ASI timeline in AI 2027, including a claim that ASI in that scenario could arrive four months later if the current pace continued. Those time references are presented as quoted assessment or scenario discussion, not as events that have already occurred.
The source article was published by the WeChat account New Intelligence and credited to the author ASI Apocalypse. Its cited reference posts were published by METR and related accounts.


