The model Anthropic will not release is already working inside the company

Anthropic's risk report for August, published in redacted form, describes an internal model it calls Model 2: somewhat more capable than Mythos 5, with no current plans for external release and, in the company's own words, not put through all of its typical suite of predeployment assessments. The same report says Mythos 5 and Model 2 are used heavily within Anthropic for coding, data generation and other agentic use cases.

It goes further in the section on automated research. The two models are used extensively for research and engineering, the report says, both interactively and through persistent agent deployments, and Claude now writes a large majority of the code merged into Anthropic's production codebases. The company's estimate of what that buys is unusually restrained: internal AI research and development is significantly faster than it would be without AI assistance, but not yet by a factor of two, and it says measurement here is difficult.

OpenAI's posture on its own unreleased model is not the same one. It said on 7 August that preliminary evaluations of Astra showed advances in agentic coding and cybersecurity strong enough that it could not rule out the Critical capability level of its Preparedness Framework, that it had paused internal activity involving Astra which did not meet stricter security requirements, and that government agencies and independent safety organisations would evaluate the model before deployment. One lab documents heavy internal use of a model it will not ship; the other says it has stopped the internal work that does not clear a new bar.

That gap is the thing Dylan Patel put his finger on. In an episode of the SemiAnalysis podcast published on Monday, and in a clip of it that has been watched some 80,000 times on X, the SemiAnalysis founder says he has heard Anthropic finished training a model it is not releasing, notes OpenAI has gone quiet on Astra, and then asks the question worth asking: whether the labs have "prevented themselves from using Mythos 2 internally to make Mythos 3 better".

The distinction matters because only one half of it is visible. A release pause is announced; an internal pause is not, and nothing outside a lab can measure it. Benchmarks, open-weight comparisons and every public account of who is ahead are built from what ships, so a lab that keeps its strongest system inside can look like it is slowing down while the frontier it works with moves on. Patel's other point follows from that: the gap to open models could appear to narrow for reasons that have nothing to do with open models catching up.

What none of this establishes is the loop itself. Anthropic does not say Model 2 is being used to train its successor, and does not say it is not; there is no system card, benchmark or independent evaluation of the model; the report is redacted and its coverage date is 15 July. Mythos 2 is the name Patel used for what he had heard, and it is not the name in Anthropic's document. No plans to release is a plan, and plans change.