Claude Opus 4.5

Training data disclosure: Vague

Level determined for 13% of benchmark cells.

• TestTest data seen2
• SeenSeen, split unclear0
• TrainTraining data seen0
• Rel.Related data seen0
• NoNot seen1
• ?Unknown19

Other declared training data

• publicly available information from the internet up to May 2025 (broad web data: may contain benchmark test sets) · split: unknown
  Evidence

  “Claude Opus 4.5 was trained on a proprietary mix of publicly available information from the internet up to May 2025, non-public data from third parties, data provided by data-labeling services and paid contractors, data from Claude users who have opted in to have their data used for training, and data generated internally at Anthropic.”

• non-public data from third parties (not yet matched to a dataset) · split: unknown
  Evidence

  “Claude Opus 4.5 was trained on a proprietary mix of publicly available information from the internet up to May 2025, non-public data from third parties, data provided by data-labeling services and paid contractors, data from Claude users who have opted in to have their data used for training, and data generated internally at Anthropic.”

• data provided by data-labeling services and paid contractors (not yet matched to a dataset) · split: unknown
  Evidence

  “Claude Opus 4.5 was trained on a proprietary mix of publicly available information from the internet up to May 2025, non-public data from third parties, data provided by data-labeling services and paid contractors, data from Claude users who have opted in to have their data used for training, and data generated internally at Anthropic.”

• data from Claude users who have opted in to have their data used for training (not yet matched to a dataset) · split: unknown
  Evidence

  “Claude Opus 4.5 was trained on a proprietary mix of publicly available information from the internet up to May 2025, non-public data from third parties, data provided by data-labeling services and paid contractors, data from Claude users who have opted in to have their data used for training, and data generated internally at Anthropic.”

• data generated internally at Anthropic (not yet matched to a dataset) · split: unknown
  Evidence

  “Claude Opus 4.5 was trained on a proprietary mix of publicly available information from the internet up to May 2025, non-public data from third parties, data provided by data-labeling services and paid contractors, data from Claude users who have opted in to have their data used for training, and data generated internally at Anthropic.”

Declared exclusions

• Dataset MMLU (Measuring Massive Multitask Language Understanding) · test portion excluded
  Evidence

  “We scanned our training corpus for exact substring matches of the evaluations we benchmark and removed documents that contain five or more”

  “This is effective for reducing direct contamination of multiple-choice questions and answers in evaluations such as MMLU or GPQA.”

• Dataset GPQA (Graduate-Level Google-Proof Q&A) · test portion excluded
  Evidence

  “We scanned our training corpus for exact substring matches of the evaluations we benchmark and removed documents that contain five or more”

  “This is effective for reducing direct contamination of multiple-choice questions and answers in evaluations such as MMLU or GPQA.”

Contamination analyses

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

• AIME 2025 · training data

  Developer found rephrased AIME questions, official solutions and model answers in the training corpus despite decontamination; the AIME 2025 score may be inflated (example answer suggesting memorisation)

  Evidence

  “Our investigation found that rephrased AIME questions, official solutions, and model-generated answers persisted in the training corpus despite our targeted efforts to remove them.”

  “However, we have some concerns that contamination may have inflated this score, as discussed in Section 2.2.”

• SWE-bench Verified · training data

  Third-party (OpenAI) elicitation study found strong contamination examples: Claude Opus 4.5 reproduced SWE-bench Verified gold-patch / task-specific details

  Evidence

  “we tasked GPT-5 with probing a GPT-5.2-Chat, Claude Opus 4.5 and Gemini 3 Flash Preview for contamination.”

  “indicating that all of them have seen at least some of the problems and solutions during training”

Benchmark cells

Classic LLM benchmarks

Multimodal LLM benchmarks

Recent LLM benchmarks

Notes

Included (point release) because its system card introduced Anthropic's documented decontamination methodology (Section 2.2: exact-substring removal, 20-gram fuzzy decontamination, canary filtering, manual inspection), which later cards reference. The card states manual inspection 'confirmed low levels of contamination for many evaluations (e.g. Humanity's Last Exam)' and that decontamination was 'significantly improved with Claude Opus 4.5'. The docs give a training data cutoff of Aug 2025 while the system card describes internet data 'up to May 2025'. Decontamination statements not shown as cells (they name no benchmark tracked here, or none at all): Terminal-Bench. Accepts image input: “Claude Opus 4.5 can understand both text (including voice dictation) and image inputs, engaging in conversation, analysis, coding, and creative tasks.” Accepts image input: “Text and images → text”

Sources

• System Card: Claude Opus 4.5 · model card · retrieved 2026-10-15
• Claude docs: Claude Opus 4.5 model page · website · retrieved 2026-10-15
• Anthropic Transparency Hub: Model Report · website · retrieved 2026-10-15
• Why we no longer evaluate SWE-bench Verified (OpenAI, February 23, 2026) third-party study· study · retrieved 2026-10-15