GPT-5.2

2025-12-11 · Publisher : OpenAI · Domain : Artificial Intelligence

OpenAI launches GPT-5.2, its most capable series of models for professional knowledge work, with Instant, Thinking and Pro versions, better at spreadsheets, presentations, code, vision and long contexts.

What changes

  • New state of the art announced on GDPval: GPT-5.2 Thinking beats or ties professionals in 70.9% of comparisons across 44 occupations.
  • More polished creation of spreadsheets and presentations; on an internal junior investment banking analyst modeling test, the score goes from 59.1% to 68.4%.
  • Improved tool calling, long context and vision; new “xhigh” reasoning effort and a compaction endpoint to extend the effective context.
  • Gradual rollout in ChatGPT (Plus, Pro, Go, Business, Enterprise); GPT-5.1 stays in legacy models for three months.

### Specifications - 400,000-token context, 128,000-token output, text and image input, text output, cutoff August 31, 2025. - API: gpt-5.2 (Thinking), gpt-5.2-chat-latest (Instant) and gpt-5.2-pro (Responses only), snapshots 2025-12-11. - Effort levels: none, low, medium, high, xhigh. ### Pricing Price per million tokens (USD), with a 90% discount on cached input. Higher than GPT-5.1, offset according to the vendor by better token efficiency. | | Input | Cached input | Output | | --- | --- | --- | --- | | gpt-5.2 / gpt-5.2-chat-latest | 1.75 | 0.175 | 14 | | gpt-5.2-pro (documentation) | 21 | not disclosed | 168 | ### Benchmarks Scores published by the vendor, GPT-5.2 Thinking versus GPT-5.1 Thinking. | Benchmark | GPT-5.2 | GPT-5.1 | | --- | --- | --- | | GDPval (wins or ties) | 70.9% | 38.8% (GPT-5) | | SWE-bench Verified | 80.0% | 76.3% | | SWE-Bench Pro (public) | 55.6% | 50.8% | | GPQA Diamond | 92.4% | 88.1% | | AIME 2025 | 100% | 94.0% | | ARC-AGI-2 (Verified) | 52.9% | 17.6% | | FrontierMath (levels 1 to 3) | 40.3% | 31.0% | ### Key takeaways GPT-5.2 targets office deliverables (spreadsheets, presentations, analyses), with a per-token cost higher than GPT-5.1. The gains are most marked on long tasks and abstraction (ARC-AGI).

Links

Tags : LLM, OpenAI, GPT, GPT-5.2, Raisonnement, Agent