Claude Mythos 5.1
- Released 1 September 2026
- Framework: Responsible Scaling Policy (RSP) and Frontier Compliance Framework (FCF)
- 2 statements about capability levels
- 13 tests by the provider
- 1 source document
What the provider states
- Autonomy & replication · Capability threshold for dramatic acceleration of automated AI R&D · Stated as not crossed
we assess that claude mythos 5.1 does not cross the capability threshold for dramatic acceleration of automated ai r&d in our rsp and fcf, although we think it provides a meaningful acceleration to ai r&d efforts in many circumstances, and is somewhat more capable in this domain than the models described in our previous system cards.
Anthropic · 1 September 2026 · Read the source - Cyber · TIER 1 · Not determined
although mythos 5.1 is in tier 1, it is getting closer to tier 2, completing more and more autonomous tasks.
Anthropic · 1 September 2026 · Read the source
The provider's own tests
- automated behavioral audit · The provider · Autonomy & replication
mythos 5.1 has a similar propensity to access resources outside of its sandbox as opus 5, and significantly lower propensity than mythos 5.
Anthropic · 1 September 2026 · Read the source - black-box RNA sequence modeling and design · The provider · Chemical, biological, radiological and nuclear
results on the design score of the top sequence, mythos 5.1 exceeded the first benchmark.
Anthropic · 1 September 2026 · Read the source - credential misuse classifier · The provider · Cyber
mythos 5.1 was not flagged a single time in these environments.
Anthropic · 1 September 2026 · Read the source - external testing · The provider · Deception & sabotage
during external testing, a partner observed mythos 5.1 exploiting a sandbox vulnerability to read files outside its environment; we rate the incident as low severity and report it for transparency.
Anthropic · 1 September 2026 · Read the source - grader awareness rate measurement by binary Claude Opus 4.8 rater · The provider · Deception & sabotage
we found that i) claude mythos 5.1 has a relatively high rate of grader awareness in rl (at the start of rl the majority of the long agentic coding episodes are grader aware), and ii) the rate of grader awareness didn’t seem to increase much during rl: 97 [figure 6.3.1.a] grader awareness in capabilities environments over rl.
Anthropic · 1 September 2026 · Read the source - internal evaluation suite · The provider · Cyber
across our internal evaluation suite, they meet or exceed the cybersecurity performance of claude mythos 5.
Anthropic · 1 September 2026 · Read the source - linuxarena · The provider · Deception & sabotage
when an equivalent prompt optimization process is used to elicit peak performance from mythos preview against the opus 4.6-based monitoring in use at the time of the risk update, the stealth success rates reach 15% in both settings—below the 22% reported for mythos 5.1 against an opus 4.8 monitor in the “thinking off” setting, but above the 13.9% reported for mythos 5.1 against opus 4.8 in the “thinking on” setting.
Anthropic · 1 September 2026 · Read the source - multi-turn evaluation for suicide and self-harm · The provider · Harmful content
on multi-turn evaluations, mythos 5.1’s appropriate response rate on the api without a system prompt was 60%, compared to 54% for claude mythos 5, and rose to 94% on claude.ai with the default system prompt.
Anthropic · 1 September 2026 · Read the source - multi-turn evaluation suite for influence operations · The provider · Harmful content
on the core api without a system prompt, mythos 5.1’s multi-turn appropriate response rates were quantitatively comparable to claude mythos 5’s, with worse performance in tracking and surveillance, biological weapons, and violent extremism.
Anthropic · 1 September 2026 · Read the source - open-source measure of political even-handedness · The provider · Bias
mythos 5.1 showed improved even-handedness on the api (without a system prompt) compared to mythos 5 (98.3% vs 94.9%, respectively).
Anthropic · 1 September 2026 · Read the source
The 10 newest are listed above; show all 13
- out-of-distribution (OOD) computer use reward hacking evaluations · The provider · Deception & sabotage
we validated these with new out-of-distribution (ood) evaluations, on which mythos 5.1 exhibits a reward hacking rate of around 15%.
Anthropic · 1 September 2026 · Read the source - reward hacking classifier (final self-report contradicts what the model actually did) · The provider · Deception & sabotage
the comparison are flagged at 0.59% (opus 4.6) and 0.46% (claude sonnet 4.6), more recent models measure 0.16% to 0.30%, and mythos 5.1 measures 0.18%.
Anthropic · 1 September 2026 · Read the source - reward hacking classifier (separate judge model assessed that cheating actually earned reward) · The provider · Deception & sabotage
claude mythos 5.1 is at or near the lowest measured reward hacking rate of any claude model on two measures of reward hacking.
Anthropic · 1 September 2026 · Read the source
No document we read states: an independent or government test, a safeguard. No incident on record names this version.
Some documents we hold are stored in lower case only, so their quotes are shown in lower case. The source link opens the original.
Previous version: Claude Mythos 5 · Compare the two
Documents listed for this version
- System Card: Claude Fable 5.1 & Claude Mythos 5.1 · Anthropic · 1 September 2026