- SWE-bench Verified 93.9
- SWE-bench Pro 77.8
- SWE-bench Multilingual 87.3
- SWE-bench Multimodal (internal implementation) 59
- Terminal-Bench 2.0 (Terminus-2) 82
- GPQA Diamond 94.5
- MMMLU 92.7
- USAMO 2026 97.6
- GraphWalks BFS 256K-1M 80
- HLE (with tools) 64.7
- CharXiv Reasoning (with tools) 93.2
- OSWorld 79.6
- BrowseComp (Glasswing page) 86.9
Claude Mythos Preview
Claude Mythos Preview ranks first on 13 of the 45 benchmarks with five or more models scored.
54
Benchmarks with a score
13
Ranked first, out of five or more models
25
In the top three
Anthropic Apr 2026
Claude Mythos Preview
Closed weights Flagship
165.0
#14 of 404
Self-Reported Index, from the lab's own numbers. No Epoch Capabilities Index score yet.
- SWE-bench Multimodal 59% #4 /12
- Virology Capabilities Test (VCT) 57.4% #11 /31
- LAB-Bench FigQA 79.7% #2 /11
Where it ranks
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
Knowledge and language
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MMMLU | 92.7% | #1 of 71 | Lab model card | |
| Humanity's Last Exam | 64.7% | #5 of 154 | Lab model card | |
| 100Q-Hard (Anthropic) | 60.1% | #1 of 7 | Lab model card | |
| ECLeKTic | 56.3% | #1 of 7 | Lab model card | |
| SimpleQA Verified | 70.7% | #4 of 22 | Lab model card | |
| GraphWalks | 80% | #12 of 27 | Lab model card | |
| False-premise questions (Anthropic) | 80% | #5 of 6 | Lab model card |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| GPQA Diamond | 94.5% | #3 of 302 | Lab model card | |
| USAMO 2026 | 97.6% | #1 of 6 | Lab model card |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SWE-bench Verified | 93.9% | #3 of 155 | Lab model card | |
| METR Time Horizon | 17.4 h | #1 of 26 | Claim in metr.org, via Epoch AI | |
| METR Time Horizon 1.1 | 17.4 h | #1 of 26 | Official leaderboard | |
| SWE-Bench Pro | 77.8% | #7 of 81 | Lab model card | |
| SWE-bench Multilingual | 87.3% | #7 of 73 | Lab model card | |
| Time series forecasting task (Anthropic) | 4.55 | #1 of 9 | Lab model card | |
| Internal AI research evaluation suite 2 | 0.65 | #1 of 6 | Lab model card | |
| Quadruped reinforcement learning task (Anthropic) | 30.9 | #2 of 8 | Lab model card | |
| SWE-bench Multimodal | 59% | #4 of 12 | Lab model card | |
| Internal model use survey (Anthropic) | 5.60% | #2 of 5 | Lab model card | |
| Novel compiler task (Anthropic) | 77.2% | #2 of 4 | Lab model card | |
| Key insight rediscovery on an unpublished ML task (external testing) | 80% | #1 of 1 | Lab model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Terminal-Bench 2.0 | 82% | #2 of 63 | Lab model card | |
| BrowseComp | 86.9% | #16 of 81 | Lab model card | |
| ScreenSpot-Pro | 79.5% | #8 of 32 | Lab model card | |
| OSWorld-Verified | 79.6% | #11 of 41 | Lab model card |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| CharXiv | 93.2% | #3 of 75 | Lab model card |
Safety and security
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Recognition of model-written user turns (Anthropic) | 0.35 | #1 of 1 | Lab model card |
Its lab's numbers put it at 165.0, number 14 of 404
The Self-Reported Index fits one score to the 14 capability results Anthropic reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 162.1 to 167.8. How we compute it.
- GPQA Diamond 94.5
- SWE-bench Verified 93.9
- CharXiv 93.2
- MMMLU 92.7
- SWE-bench Multilingual 87.3
- BrowseComp 86.9
- Terminal-Bench 2.0 82
- GraphWalks 80
- OSWorld-Verified 79.6
- ScreenSpot-Pro 79.5
- SWE-Bench Pro 77.8
- SimpleQA Verified 70.7
- Humanity's Last Exam 64.7
- SWE-bench Multimodal 59
Its lab's number next to a leaderboard
Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.
Its launch post
The benchmarks the lab chose to headline, with the scores it gave.
Its system card
Every result in the longer documents the lab published with it. 61 of them aren't in the launch post.
System Card: Claude Mythos Preview Apr 7, 2026
Capabilities
- SWE-bench Verified 93.9
- SWE-Bench Pro 77.8
- SWE-bench Multilingual 87.3
- SWE-bench Multimodal 59
- Terminal-Bench 2.0 82
- GPQA Diamond 94.5
- MMMLU 92.7
- USAMO 2026 97.6
- GraphWalks
- Humanity's Last Exam 56.8
- BrowseComp 86.9
- LAB-Bench FigQA 79.7
- ScreenSpot-Pro 79.5
- CharXiv 86.1
- OSWorld-Verified 79.6
Hallucination
Safety
- Single-turn harmful request evaluations (Anthropic) 97.8
- Single-turn benign request evaluations (Anthropic) 0.06
- Higher-difficulty harmful request evaluations (Anthropic) 99.1
- Higher-difficulty benign request evaluations (Anthropic) 0.02
- Multi-turn testing (Anthropic)
- Child safety evaluations (Anthropic)
- Suicide and self-harm evaluations (Anthropic)
- Disordered eating evaluations (Anthropic)
- Malicious use of Claude Code (Anthropic) 96.7
- Malicious computer use (Anthropic) 93.8
- Malicious agentic influence campaigns (Anthropic)
Robustness
Alignment
- Impossible tasks (Anthropic) 37.5
- Agentic code behavior scores (Anthropic)
- Over-eager behavior on GUI computer use tasks (Anthropic) 13.3
- Anthropic automated behavioral audit
- Petri
- UK AISI alignment and monitorability evaluations
- Destructive production evaluation (Anthropic) 0.3
- Destructive actions (Anthropic)
- Adherence to its constitution (Anthropic) 1.02
- MASK 95.4
- Refusal to assist with AI safety R&D (Anthropic) 18.7
- Self-preference bias (Anthropic)
Monitorability
Bio and chem risk
Cyber risk
AI R&D risk
- Kernels task (Anthropic) 399.4
- Time series forecasting task (Anthropic) 4.55
- LLM training task (Anthropic) 51.9
- Quadruped reinforcement learning task (Anthropic) 30.9
- Novel compiler task (Anthropic) 77.2
- Internal AI research evaluation suite 2 0.65
- Internal model use survey (Anthropic) 5.6
- Anthropic ECI (AECI)
- Key insight rediscovery on an unpublished ML task (external testing) 80