Anthropic Unreleased AI Models: METR Finds Three in Six Months
METR found three Anthropic unreleased AI models in six months. Learn what the timeline says about AI-assisted research - and what it does not prove.

Table of Contents
Quick Answer
METR reported that Anthropic trained three unreleased AI models during a six-month period, with training dates spanning July to October. The report also assessed that Anthropic's model capabilities on well-specified tasks were improving at around twice the rate observed between March 2024 and October 2025, but it did not show that AI systems autonomously built their successors.
The finding adds detail to an important question in AI: how quickly are leading labs improving their models, and how much are those models helping to advance the next generation?
What METR reported about Anthropic's unreleased AI models
METR identified three models by codenames shared with its researchers: Model 2, trained in July; Model 3, trained between August and September; and Model 4, trained between September and October. The report cautions that these codenames may not match Anthropic's internal names.
The models were not publicly released. Their training dates therefore provide a view of Anthropic's development activity that is separate from the schedule of products customers can use. METR also identified an earlier unreleased model, called Unreleased 2022, dating to 2022.
The findings are based on METR's assessment of Anthropic's model-development history. They offer evidence of a brisk training cadence, but the codenames and dates alone do not establish what the models can do or whether each represents a major capability jump.
Why METR says Anthropic's AI development may be accelerating
METR's broader assessment is that Anthropic was likely improving model capabilities on well-specified tasks at around twice the rate it did between March 2024 and October 2025. The researchers said this acceleration likely involved extensive use of AI in research and development - including AI tools assisting human researchers.
That distinction matters. AI systems can make research more efficient by helping with tasks such as coding, analysis, or experiments, while people continue to set goals, judge results, and direct the work. More capable assistance could shorten development cycles even if AI is not independently designing and training new models.
METR said it saw no clear signs of diminishing returns or observable bottlenecks that would indicate this pace is about to level off. But the researchers also warned that fully autonomous contributions could play a larger role in the future. Their assessment describes observed development and a possible trajectory, not a guarantee that progress will continue at the same rate.
What the findings do - and do not - say about self-improving AI
Three training runs in six months are evidence of development activity, not proof that Anthropic's models are independently building their successors. METR's account points primarily to AI assisting human researchers. The report's concern is that this role could expand, potentially making AI research and development faster and less dependent on direct human work.
Anthropic co-founder Sam McCandlish said the company planned to share its own analysis of the speed of AI progress within Anthropic. Separately, Anthropic had said it had not designated any current model as exceeding the capability threshold in its Responsible Scaling Policy for dramatically accelerated progress. That statement is not the same as a rebuttal to METR's historical analysis: one concerns Anthropic's policy classification of current models, while the other assesses the pace of past development.
METR's report is best read as a signal to watch the role of AI in AI research, as well as the number of models being trained. The training timeline suggests a quick cadence; the larger question is whether AI assistance will become autonomous enough to materially drive future progress.
Follow our AI research coverage for updates on METR's findings and Anthropic's forthcoming analysis of AI progress.
Frequently Asked Questions
What did METR find about Anthropic's unreleased AI models?
Does the METR report prove Anthropic's AI models are self-improving?
How quickly did Anthropic's AI capabilities improve, according to METR?
Why are Anthropic's unreleased models important?
Key Takeaways
- METR identified three unreleased Anthropic models, labeled Model 2, Model 3, and Model 4, with training dates from July through October.
- The report's codenames may not match Anthropic's internal model names, and the training dates do not reveal each model's capabilities.
- METR assessed that capability improvements on well-specified tasks were happening at around twice the prior rate.
- METR attributed the acceleration likely in part to AI tools assisting human researchers, not to fully autonomous AI research.
- The report describes observed development and a possible trajectory, not proof that the pace will continue or that AI is independently designing successor models.