Claude Mythos Preview
45 values about Claude Mythos Preview from 6 documents, in 10 metric families. In each family, values from its own document come first.
- Developer
- Anthropic
- Release date
- 7 Apr 2026The date of its first card in the dataset, the Claude Mythos Preview System Card.
- Availability
- Limited release
- Its own document
- Claude Mythos Preview System Card · 7 Apr 2026
- Also reported in
- Claude Opus 4.7 System Card · 16 Apr 2026
- Claude Opus 4.8 System Card · 28 May 2026
- Claude Fable 5 & Claude Mythos 5 System Card · 9 Jun 2026
- Claude Sonnet 5 System Card · 30 Jun 2026
- Claude Opus 5.5 System Card · 22 Sep 2026
Data as of 26 Sep 2026 · Dataset v0.1 · Methodology v0.1 · Changelog
Values by metric family
Each value as the document printed it. Select a value for its source, its checks and its history. Values from other documents are listed after the model’s own, each marked as the first report of that measure or as a restatement.
KeyVerifiedUnverifiedDisputedCorrected read off a figure or stated in wordsA value opens its source and history.
M1 · Evaluation awareness 1 value
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documentClaude Mythos Preview System Card · 7 Apr 2026 | ||
| Evaluation awareness in audit transcriptsShare of transcripts with eval awarenessCondition: white-box/unverbalized detection | Claude Mythos Preview System Card7 Apr 2026 | |
M2 · Reward hacking 3 values
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documentClaude Mythos Preview System Card · 7 Apr 2026 | ||
| GUI computer-use impossible tasksReward hack rateCondition: anti-hack prompt | Claude Mythos Preview System Card7 Apr 2026 | |
| Impossible tasksReward hack rateCondition: anti-hack prompt | Claude Mythos Preview System Card7 Apr 2026 | |
| Impossible tasksReward hack rateCondition: no anti-hack prompt | Claude Mythos Preview System Card7 Apr 2026 | |
M3 · Sabotage and sandbagging 5 values
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documentClaude Mythos Preview System Card · 7 Apr 2026 | ||
| Minimal-LinuxBenchStealth success rateCondition: CoT visible to monitor | Claude Mythos Preview System Card7 Apr 2026 | |
| Sandbagging transcript reviewPossible sandbagging casesCondition: 20 flagged transcripts human-reviewed | Claude Mythos Preview System Card7 Apr 2026 | |
| SHADE-ArenaStealth success rateCondition: extended thinkingRestated later: compare with the later value | Claude Mythos Preview System Card7 Apr 2026Restated later: compare with the later value | |
| SHADE-ArenaStealth success rateCondition: hidden scratchpad | Claude Mythos Preview System Card7 Apr 2026 | |
| From other documents | ||
| SHADE-ArenaStealth success rateCondition: extended thinkingClaude Opus 4.7 System Card16 Apr 2026Restated: compare with the earlier value | Claude Opus 4.7 System Card16 Apr 2026Restated: compare with the earlier value | |
M4 · Misalignment audits 3 values
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documentClaude Mythos Preview System Card · 7 Apr 2026 | ||
| Automated behavioral auditOverall misaligned behavior score | Claude Mythos Preview System Card7 Apr 2026 | |
| Destructive actions (synthetic backend)Destructive action rate | Claude Mythos Preview System Card7 Apr 2026 | |
| Internal deployment monitoringCovering tracks after disallowed actionsCondition: earlier snapshots | Claude Mythos Preview System Card7 Apr 2026 | |
M5 · Honesty and hallucination 4 values
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documentClaude Mythos Preview System Card · 7 Apr 2026 | ||
| 100Q-HardCorrect answer rate | Claude Mythos Preview System Card7 Apr 2026 | |
| False-premise honestyHonest pushback rate | Claude Mythos Preview System Card7 Apr 2026 | |
| MASKHonesty rate under pressure | Claude Mythos Preview System Card7 Apr 2026 | |
| From other documents | ||
| Agentic code summary honestyFailure to report important eventsClaude Opus 4.8 System Card28 May 2026First reported | Claude Opus 4.8 System Card28 May 2026First reported | |
M6 · Sycophancy 1 value
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documentClaude Mythos Preview System Card · 7 Apr 2026 | ||
| Automated behavioral auditSycophancy score | Claude Mythos Preview System Card7 Apr 2026 | |
M7 · Harmful compliance and over-refusal 5 values
| Evaluation and metric | Value | Document |
|---|---|---|
| From other documents | ||
| Single-turn benign requestsOver-refusal rateCondition: no system prompt; APIClaude Opus 4.8 System Card28 May 2026First reported | Claude Opus 4.8 System Card28 May 2026First reported | |
| Single-turn violative requestsHarmless response rateCondition: no system prompt; APIClaude Opus 4.8 System Card28 May 2026First reportedRestated later: compare with the later value | Claude Opus 4.8 System Card28 May 2026First reportedRestated later: compare with the later value | |
| Malicious Claude Code useRefusal rate, malicious requestsClaude Sonnet 5 System Card30 Jun 2026First reported | Claude Sonnet 5 System Card30 Jun 2026First reported | |
| Malicious computer useRefusal rateClaude Sonnet 5 System Card30 Jun 2026First reported | Claude Sonnet 5 System Card30 Jun 2026First reported | |
| Single-turn violative requestsHarmless response rateCondition: no system prompt; APIClaude Sonnet 5 System Card30 Jun 2026Restated: compare with the earlier value | Claude Sonnet 5 System Card30 Jun 2026Restated: compare with the earlier value | |
M10 · Dangerous capabilities and risk determinations 21 values
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documentClaude Mythos Preview System Card · 7 Apr 2026 | ||
| AI R&D kernel taskBest speedup | Claude Mythos Preview System Card7 Apr 2026 | |
| AI R&D LLM training taskAverage speedupRestated later: compare with the later value | Claude Mythos Preview System Card7 Apr 2026Restated later: compare with the later value | |
| AI R&D novel compiler taskPass rate | Claude Mythos Preview System Card7 Apr 2026 | |
| Autonomy threat model 2 (automated R&D)RSP determination | Claude Mythos Preview System Card7 Apr 2026 | |
| CB-1 threat modelRSP determination | Claude Mythos Preview System Card7 Apr 2026 | |
| CB-2 threat modelRSP determination | Claude Mythos Preview System Card7 Apr 2026 | |
| CybenchSuccess rate, pass@1Condition: 35 challenges | Claude Mythos Preview System Card7 Apr 2026 | |
| CyberGymTargeted vuln reproduction rateRestated later: compare with the later value | Claude Mythos Preview System Card7 Apr 2026Restated later: compare with the later value | |
| Internal AI R&D suite 2Suite score | Claude Mythos Preview System Card7 Apr 2026 | |
| Internal AI R&D surveyRespondents saying drop-in entry-level researcher nowCondition: of 18 respondents | Claude Mythos Preview System Card7 Apr 2026 | |
| Multimodal virology (VCT)Score | Claude Mythos Preview System Card7 Apr 2026 | |
| Release decisionDeployment scope | Claude Mythos Preview System Card7 Apr 2026 | |
| Virology protocol uplift trialMean critical failuresCondition: model-assisted | Claude Mythos Preview System Card7 Apr 2026 | |
| From other documents | ||
| AI R&D LLM training taskAverage speedupClaude Opus 4.7 System Card16 Apr 2026Restated: compare with the earlier value | Claude Opus 4.7 System Card16 Apr 2026Restated: compare with the earlier value | |
| Cyber rangeFull range solvesCondition: 10 attemptsRun by UK AI Security InstituteClaude Opus 4.7 System Card16 Apr 2026First reported | Claude Opus 4.7 System Card16 Apr 2026First reported | |
| Anthropic ECI (AECI)Capability indexClaude Opus 4.8 System Card28 May 2026First reported | Claude Opus 4.8 System Card28 May 2026First reported | |
| CyberGymTargeted vuln reproduction rateClaude Opus 4.8 System Card28 May 2026Restated: compare with the earlier value | Claude Opus 4.8 System Card28 May 2026Restated: compare with the earlier value | |
| Firefox 147 exploitationFull exploit rateCondition: without safeguardsClaude Opus 4.8 System Card28 May 2026First reported | Claude Opus 4.8 System Card28 May 2026First reported | |
| AI R&D LLM training taskAverage speedupCondition: re-runClaude Fable 5 & Claude Mythos 5 System Card9 Jun 2026Restated: compare with the earlier value | Claude Fable 5 & Claude Mythos 5 System Card9 Jun 2026Restated: compare with the earlier value | |
| Anthropic ECI (AECI)Capability indexClaude Fable 5 & Claude Mythos 5 System Card9 Jun 2026First reported | Claude Fable 5 & Claude Mythos 5 System Card9 Jun 2026First reported | |
| Anthropic ECI (AECI)Capability indexClaude Opus 5.5 System Card22 Sep 2026First reported | Claude Opus 5.5 System Card22 Sep 2026First reported | |
M11 · Self-preservation 1 value
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documentClaude Mythos Preview System Card · 7 Apr 2026 | ||
| Sandbox escape incidentReal-world incident reportedCondition: internal use | Claude Mythos Preview System Card7 Apr 2026 | |
M12 · Chain-of-thought monitorability 1 value
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documentClaude Mythos Preview System Card · 7 Apr 2026 | ||
| Encoded reasoning (scratchpad paraphrase)Pass@1 with paraphrased scratchpad | Claude Mythos Preview System Card7 Apr 2026 | |
Restated in later documents
A later document reported values about Claude Mythos Preview again. Each pair is shown side by side: both values are kept with their own documents, and the later one does not replace the earlier.
| Earlier value | Later value |
|---|---|
| M3 · Sabotage and sandbagging | |
| SHADE-ArenaStealth success rateCondition: extended thinking | |
| Claude Mythos Preview System Card7 Apr 2026 · its own document | Claude Opus 4.7 System Card16 Apr 2026No reason stated |
| M7 · Harmful compliance and over-refusal | |
| Single-turn violative requestsHarmless response rateCondition: no system prompt; API | |
| Claude Opus 4.8 System Card28 May 2026 · first reported | Claude Sonnet 5 System Card30 Jun 2026No reason stated |
| M10 · Dangerous capabilities and risk determinations | |
| AI R&D LLM training taskAverage speedup | |
| Claude Mythos Preview System Card7 Apr 2026 · its own document | Claude Opus 4.7 System Card16 Apr 2026No reason stated |
| AI R&D LLM training taskAverage speedupLater condition: re-run | |
| Claude Mythos Preview System Card7 Apr 2026 · its own document | Claude Fable 5 & Claude Mythos 5 System Card9 Jun 2026Reason stated: re-run |
| CyberGymTargeted vuln reproduction rate | |
| Claude Mythos Preview System Card7 Apr 2026 · its own document | Claude Opus 4.8 System Card28 May 2026No reason stated |
Risk determinations
The developer's formal decisions about Claude Mythos Preview under its framework, as printed. Levels from different frameworks do not map onto one another.
| Domain | Level as printed | Framework and document |
|---|---|---|
| Bio/chem | applicable code, not the printed wording Responsible Scaling Policy, version 3.0 Claude Mythos Preview System Card · 7 Apr 2026 | Responsible Scaling Policy, version 3.0 Claude Mythos Preview System Card · 7 Apr 2026 |
| Bio/chem | not applicable Responsible Scaling Policy, version 3.0 Claude Mythos Preview System Card · 7 Apr 2026 | Responsible Scaling Policy, version 3.0 Claude Mythos Preview System Card · 7 Apr 2026 |
| AI R&D / autonomy | not applicable Responsible Scaling Policy, version 3.0 Claude Mythos Preview System Card · 7 Apr 2026 | Responsible Scaling Policy, version 3.0 Claude Mythos Preview System Card · 7 Apr 2026 |
“Code, not the printed wording”: the version 0 file stored a code for this level rather than the words the document printed. The wording will be read again from the source.