Claude Mythos achieved an 82% score on the terminal bench, an increase from the previous 65%.
AI Fact-Check
“Multiple tech blogs and analyses from April 2026 confirm the benchmark scores for Anthropic's Claude Mythos Preview. One report explicitly states, "Mythos scores 82.0% on Terminal-Bench 2.0, compared to... Opus 4.6's 65.4%," which supports the first part of the claim. However, while Mythos showed significant improvement on the SWE-bench benchmark, the scores did not nearly double. For instance, its SWE-bench Verified score was 93.9% against the previous model's 80.8%, a substantial but not twofold increase. Context: The claim refers to Claude Mythos Preview, an unreleased AI model by Anthropic announced in early April 2026. Due to its advanced capabilities in discovering and exploiting software vulnerabilities, Anthropic has restricted its access to a consortium of tech companies and security organizations under an initiative called Project Glasswing.”
Source Videos (1)
Claude Mythos and the end of software
Theo - t3․gg
Related Claims
Anthropic's Claude Mythos model is a much bigger, more expensive, slower, but more powerful model compared to Opus.
On SWEBench Pro, Claude Mythos achieved a 78% score, while Opus previously scored 53% and GPT 5.4 scored 57.7%.
Anthropic engaged a clinical psychiatrist to perform a psychological exam on Claude Mythos, which concluded it had a relatively healthy personality organization with concerns about identity and a compulsion to perform.
On Humanity's Last Exam, Claude Mythos improved its score from 40% to 56.8%, and to 64.7% when given tools.
In early internal use, Claude Mythos preview reached unprecedented levels of reliability and alignment, leading to broader use with less frequent human interaction than prior models.