GPT-5.2

Training data disclosure: Partial list

Level determined for 4% of benchmark cells.

• TestTest data seen1
• SeenSeen, split unclear0
• TrainTraining data seen0
• Rel.Related data seen0
• NoNot seen0
• ?Unknown21

Other declared training data

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

  “The training data for GPT-5.2 includes text from Common Crawl.”

• Other publicly available datasets (broad, multi-domain text and image datasets from public repositories, online platforms and specialized websites) (broad web data: may contain benchmark test sets) · split: unknown
  Evidence

  “Other publicly available datasets include broad, multi-domain text and image datasets made available by third parties through public repositories, online platforms, and specialized websites”

• Web content crawled by GPTBot (approximately 2018 - August 2025) (broad web data: may contain benchmark test sets) · split: unknown
  Evidence

  “GPTBot is used to crawl content that may be used to train OpenAI's generative AI foundation models”

• Synthetic data generated with GPT-5, o-series and GPT-4 family models (not yet matched to a dataset) · split: unknown
  Evidence

  “OpenAI generated synthetic data using its own general-purpose AI models, including GPT-5, o-series models, and GPT-4 family models.”

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 analysis found GPT-5.2 (and GPT-5.2-Chat) could reproduce SWE-bench Verified gold patches / task details and solved 31 tasks judged almost impossible without prior exposure; OpenAI stopped reporting SWE-bench Verified

  Evidence

  “We first encountered signs of contamination in our own models. For example, when GPT-5.2 solved 31 tasks we identified to be almost impossible to solve.”

  “Given a short snippet from the task description, GPT-5.2 outputs the exact gold patch.”

  “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 as a separate release because OpenAI's EU summary says GPT-5.2 is not a modification or fine-tune of a prior model, and because OpenAI published a contamination finding for it. The GPT-5.2 system card update could not be fetched (openai.com returns 403). Other GPT-5.x releases (5.1, 5.4, 5.5, 5.6) are not covered. Accepts image input: “Modalities Text Input and output Image Input only”

Sources

• Public Summary of Training Content for GPT-5.2 (EU AI Act), v1, 30 July 2026 · website · retrieved 2026-10-15
• OpenAI API docs: GPT-5.2 model page · website · retrieved 2026-10-15
• Why we no longer evaluate SWE-bench Verified (OpenAI, February 23, 2026) · website · retrieved 2026-10-15