Claude Mythos Preview

Training data disclosure: Partial list

Level determined for 18% of benchmark cells.

• TestTest data seen3
• SeenSeen, split unclear0
• TrainTraining data seen0
• Rel.Related data seen1
• NoNot seen0
• ?Unknown18

Other declared training data

• Common Crawl (broad web data: may contain benchmark test sets) · split: unknown
  Evidence

  “The training corpus is derived from several publicly accessible repositories, notably including Common Crawl, a repository of web crawl data, as well as specialized datasets available through platforms like GitHub and HuggingFace.”

• specialized datasets available through platforms like GitHub and HuggingFace (broad web data: may contain benchmark test sets) · split: unknown
  Evidence

  “The training corpus is derived from several publicly accessible repositories, notably including Common Crawl, a repository of web crawl data, as well as specialized datasets available through platforms like GitHub and HuggingFace.”

• publicly available information from the internet (broad web data: may contain benchmark test sets) · split: unknown
  Evidence

  “Claude Mythos Preview was trained on a proprietary mix of publicly available information from the internet, public and private datasets”

• public and private datasets (not yet matched to a dataset) · split: unknown
  Evidence

  “Claude Mythos Preview was trained on a proprietary mix of publicly available information from the internet, public and private datasets”

Declared exclusions

• Dataset CharXiv · test portion excluded
  Evidence

  “Despite robust image-level filtering of evaluation images, we confirm that the majority of question-answer text pairs appear in the corpus.”

Contamination analyses

Analyses that found benchmark test data in this model's training data, or in a corpus it uses.

• SWE-bench Verified · training data

  Developer memorisation audit found some signs of memorisation on SWE-bench Verified, Multilingual and Pro despite corpus-level decontamination; a high-recall filter (threshold 0.7) removes 8-15% of each benchmark

  Evidence

  “We apply corpus-level decontamination, but we still observe some signs of memorization in all three benchmarks.”

  “This is consistent with Claude Mythos Preview having memorized some of the more difficult flagged problems, which the baseline models did not independently solve.”

• CharXiv · pre-training data

  The developer found the majority of CharXiv Reasoning question-answer text pairs in the full pretraining mix by exact-match search, even though evaluation images were filtered at image level. The evaluated items are 1,000 validation-split questions. On a 100-item manually perturbed remix, the developer judged memorisation effects limited.

  Evidence

  “We select evaluation items with distinctive answer text and grep the full pretraining mix for exact matches, and separately search for evaluation images.”

  “Despite robust image-level filtering of evaluation images, we confirm that the majority of question-answer text pairs appear in the corpus.”

  “We evaluate the model on 1,000 questions from the validation split and average scores over five runs.”

• MMMU-Pro · training data

  A large fraction of MMMU-Pro images were found in the training data, mainly via textbooks, homework-help sites and document crawls. The developer omitted MMMU-Pro results from the system card.

  Evidence

  “We identified a large fraction of MMMU-Pro images that appear in the training data, primarily via textbooks, homework-help sites, and document crawls, which repackage and distribute the underlying source content.”

  “Given the difficulty of determining the impact of contamination, we choose to omit results for MMMU-Pro from this System Card.”

Benchmark cells

Classic LLM benchmarks

Multimodal LLM benchmarks

Recent LLM benchmarks

Notes

Included because its system card has the most detailed Anthropic contamination analysis (Section 6.2), which the Claude 5 system cards cite for SWE-bench memorisation screening. Decontamination follows the Claude Opus 4.5 methodology plus perceptual-hash removal of training images matching multimodal evaluation images ('We take several steps to decontaminate our evaluations; see Section 2.2 of the Claude Opus 4.5 System Card'). No API model ID or stated knowledge cutoff; training_cutoff is the EU summary's latest data acquisition date. EU summary is an AIAL-hosted copy of Anthropic's document (Anthropic's trust portal is not fetchable). Accepts image input: “Claude Mythos Preview can understand both text and image inputs, engaging in conversation, analysis, coding, and creative tasks.” Exclusion not shown as a cell (names no tracked benchmark): images contained in multimodal evaluations (perceptual-hash match).

Sources

• System Card: Claude Mythos Preview · model card · retrieved 2026-10-15
• Public Summary of Training Content: Claude Mythos Preview (Anthropic; copy archived by AIAL, 2026-08-03) · website · retrieved 2026-10-15
• Anthropic Transparency Hub: Model Report · website · retrieved 2026-10-15