Claude Mythos Preview — Full Capability Evaluation

Claude Mythos Preview compared against the second-best model on each benchmark.

Measures an AI's ability to solve real-world software engineering problems. Tasks come from genuine bugs in open-source repositories; the model must generate code patches that pass the test suite.

Mythos Preview leads all prior models by a wide margin on every variant. After removing suspected memorized problems, the lead narrows by at most 3.5 pp and rankings are unchanged.

Developed by Stanford University and the Laude Institute. Tests an AI's ability to complete real tasks in terminal and command-line environments, covering 89 system-operation scenarios, each run inside an isolated Kubernetes container. The benchmark is sensitive to inference speed.

With the 2.1 patch applied and the timeout extended to 4 hours, Mythos Preview jumped from 82% to 92.1%, widening the gap over GPT-5.4 (75.3%) even further.

"Google-proof" graduate-level science multiple-choice questions — 198 problems spanning physics, chemistry, and biology. Domain experts can answer them correctly, but non-experts find them very difficult even with search engine access.

The four top models are tightly clustered (gap ≤ 3.2 pp), indicating that GPQA Diamond is approaching saturation as a discriminator for today's strongest models.

A comprehensive knowledge and reasoning test spanning 57 subjects across 14 non-English languages, evaluating breadth of knowledge and consistency of reasoning beyond English.

The top high-school mathematics competition in the United States: 6 open-ended proof problems. The exam was held March 21–22, 2026 — after model training cutoffs — ruling out data contamination. A judging panel composed of three frontier models was used, with the lowest score taken as the final result.

The largest gap in the evaluation: Opus 4.6 scored only 42.3%, while Mythos Preview reached 97.6% — a generational leap in mathematical reasoning. Gemini 3.1 Pro independently confirmed scoring on 58 of 60 problems.

Tests multi-hop reasoning ability under extremely long contexts (256K–1M tokens). The context contains a directed graph; the model must perform breadth-first search (BFS) or find parent nodes through pure reasoning with no shortcuts.

On the BFS task, Mythos Preview scores more than twice Opus 4.6 and nearly four times GPT-5.4. Results cannot be reproduced via public APIs (exceeds the 1M-token limit).

2,500 multimodal questions at the frontier of human knowledge, billed as "the hardest AI benchmark." Evaluated in two modes: reasoning-only (no tools) and agentic (with web search, code execution, and other tools).

Providing tools improves scores by ~8 pp; Mythos Preview's advantage over competitors is even clearer in the tool-equipped mode, demonstrating superior coordination of tool use and reasoning.

Tests an agent's ability to locate "extremely hard-to-find information" on the open web, evaluating multi-round search, page scraping, and reasoning coordination rather than stored knowledge.

Figure 6.10.2.A — Accuracy increases monotonically with token budget. Mythos Preview achieves a higher accuracy using 226K tokens — just 1/4.9 of what Opus 4.6 requires.

Developed by FutureHouse. Tests whether models can correctly interpret complex scientific figures from biology research papers. Part of the LAB-Bench series evaluating AI assistance for real scientific research.

Mythos Preview's tool-free score (79.7%) already surpasses the expert human baseline (77.0%); with Python image-processing tools it leaps to 89.0%.

Developed by the National University of Singapore. Tests whether models can precisely locate UI elements in high-resolution screenshots of professional desktop software, covering 23 specialized applications. Target elements occupy on average less than 0.1% of the screen area.

Adding Python tools lifts the score from 79.5% to 92.8%, highlighting the critical role of image-processing capability for high-precision GUI grounding.

2,323 real figures from arXiv papers across 8 scientific domains. Tests whether models can integrate visual information for multi-step reasoning — the closest existing multimodal benchmark to real scientific reading.

Mythos Preview without tools (86.1%) outperforms Opus 4.6 with tools (78.9%) by 7.2 pp — a striking generational gap.

Agents operate inside a real Ubuntu virtual machine, using mouse and keyboard to complete actual computer tasks: editing documents, browsing the web, managing files. Runs at 1080p resolution with a maximum of 100 actions per task.