Claude Mythos Preview posts leaner, stronger runs

The clean read is that Mythos looks like a larger, more efficient frontier model whose gains likely come from a mix of scale, better data, and tighter test-time strategy. Token efficiency is interesting, but it is not a clean proxy for “better pretraining.”

• –Anthropic reports large gaps on coding and agent benchmarks, including SWE-bench, Terminal-Bench, CyberGym, and BrowseComp
• –Lower token usage can mean better reasoning efficiency, but it can also reflect different budget policies, prompting, or tool-use behavior
• –Anthropic itself flags possible memorization on Humanity’s Last Exam, so the benchmark story needs caveats
• –If these numbers hold up, task-level cost may still fall even if per-token pricing rises, which matters for long-running agent workflows
• –The biggest implication is not “pretraining is solved,” but that frontier performance may be shifting toward models that spend tokens more selectively