Anthropic keeps a Claude model called Model 2 entirely to itself, and by the company’s own account, it is the most capable system Anthropic has ever built. The company disclosed this not through a launch but inside its own Risk Report, published in August 2026, which means the admission arrived with no product attached and no timeline for one. That gap between what Anthropic keeps internally and what it sells to customers says as much about its safety process as about its ambitions.
Maximilian Schreiner reported the figures for The Decoder on August 20, drawing on the Risk Report Anthropic released earlier that month. The report identifies the system as Model 2, locating it inside Anthropic’s Mythos family, the same class as the public Claude Mythos 5. Anthropic describes Model 2 as marginally stronger overall than Mythos 5, though it lags on certain individual tasks. On Anthropic’s internal AECI capability index, the model scores roughly 1.5 points higher than Mythos 5, a smaller jump than the one Anthropic recorded between Mythos Preview and Mythos 5. Anthropic did not add Model 2 to the AECI chart it publishes alongside the report, which leaves outsiders no way to place it against competing systems.
Anthropic put Model 2 through an internal safety review before deployment, though the testing was less extensive than what Mythos 5 underwent ahead of its public release. The company reports no newly discovered misalignments during that process and scores the overall misalignment risk as “low.” Anthropic has no current plans to release Model 2 to the public.
Inside the company, Model 2 is already load-bearing. Anthropic relies on it for coding, for generating synthetic data, and for supporting research and engineering work, in some cases running it through agents that operate continuously without step-by-step human supervision. Claude, according to the report, now handles most of the code that runs Anthropic’s own production systems. That detail matters more than the benchmark score: a lab running its unreleased frontier model at that scale is effectively its own toughest customer, and the “low” misalignment rating comes out of that same internal loop, not an outside audit.
The Decoder notes that Anthropic is the only source behind the AECI score, the misalignment finding, and the claim about who writes its production code. No outside lab has verified the 1.5-point gap or examined Model 2 directly.
Operators benchmarking against Claude Mythos 5 should treat that score as the floor of Anthropic’s real capability, not the ceiling: the company is already running something stronger behind closed doors, and nothing in the Risk Report says when, or whether, that gap closes for paying customers.
The Decoder’s Maximilian Schreiner reported this story on August 20, 2026, citing Anthropic’s own Risk Report from August 2026.
