AI/TLDR

GPT-5.2

OpenAI's flagship reasoning model for professional knowledge work and long-running agents.

Overview

GPT-5.2 is OpenAI's flagship model released on December 11, 2025, positioned as its most capable series for professional knowledge work and long-running agents. It came out shortly after a reported internal "code red" at OpenAI to respond to competitive pressure from Google's Gemini 3, and OpenAI says GPT-5.2 reclaims the lead on a broad set of benchmarks spanning coding, math, reasoning, and multimodal tasks.

GPT-5.2 ships in three tiers. GPT-5.2 Instant (API id gpt-5.2-chat-latest) is the fast everyday model; GPT-5.2 Thinking (API id gpt-5.2) is the reasoning model aimed at complex multi-step work, coding, and agents; and GPT-5.2 Pro (API id gpt-5.2-pro) targets the hardest technical and analytical problems with the most inference-time compute. All three accept text and image input and return text, with a 400K-token context window and up to 128K output tokens.

The headline result is GDPval, OpenAI's eval of well-specified knowledge-work tasks across 44 occupations: GPT-5.2 Thinking beats or ties top industry professionals on 70.9% of comparisons judged by human experts, which OpenAI frames as the first model to perform at or above human-expert level on that benchmark. GPT-5.2 also reaches 100% on AIME 2025 (no tools), 80.0% on SWE-bench Verified, and 52.9% on ARC-AGI-2, with a knowledge cutoff of August 31, 2025.

Released2025-12-11
LicenseProprietary
WeightsAPI only
ParametersNot disclosed by OpenAI
Context400K
Max output128K
ArchitectureProprietary; OpenAI has not disclosed the parameter count or architecture details. GPT-5.2 ships in three tiers — Instant, Thinking, and Pro — sharing one model family, where Thinking and Pro spend additional inference-time compute on step-by-step reasoning.
Knowledge cutoffAugust 31, 2025
ModalitiesText, Vision
StatusAvailable

Benchmarks

  1. GDPval (win/tie vs. experts, Thinking)70.9%
  2. AIME 2025 (no tools, Thinking)100%
  3. SWE-bench Verified (Thinking)80%
  4. SWE-bench Pro (Thinking)55.6%
  5. GPQA Diamond (Thinking)92.4%
  6. FrontierMath Tier 1-3 (Python tools, Thinking)40.3%
  7. ARC-AGI-2 (Thinking)52.9%
  8. ARC-AGI-1 (Thinking)86.2%
  9. Tau2-bench Telecom (Thinking)98.7%
  10. MMMU-Pro (vision)86.5%
  11. Video-MMMU (vision)90.5%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

Pricing

Input$1.75 / 1M tokens per 1M tokens
Cached input$0.175 / 1M tokens per 1M tokens
Output$14.00 / 1M tokens per 1M tokens

Prices are for GPT-5.2 Thinking (API id gpt-5.2). GPT-5.2 Pro is priced higher at $21 input / $168 output per 1M tokens.

Pricing source ↗

Strengths

  • First model OpenAI reports at or above human-expert level on GDPval knowledge-work tasks (70.9% win/tie rate, Thinking)
  • Strong agentic and tool-use performance, with 98.7% on Tau2-bench Telecom and a 400K-token context window for long-running, multi-step work
  • Top-tier reasoning: 100% on AIME 2025 (no tools) and a leading 52.9% on ARC-AGI-2, well ahead of competing frontier models at launch
  • Competitive coding ability at 80.0% on SWE-bench Verified and 55.6% on SWE-bench Pro
  • Improved multimodal/vision understanding, including 86.5% on MMMU-Pro and 90.5% on Video-MMMU for chart, document, and UI reasoning

Best for

  • Long-running autonomous agents and multi-step workflows that need a large context window and reliable tool use
  • Software engineering: writing, refactoring, and fixing code, and resolving real GitHub issues
  • Professional knowledge work — building spreadsheets, presentations, analysis, and reports across many occupations
  • Hard math, science, and quantitative reasoning where the Thinking and Pro tiers add inference-time compute
  • Document, chart, and UI understanding through its text-plus-vision multimodal input

How to access

ProviderModel ID
OpenAI ↗gpt-5.2
OpenAI ↗gpt-5.2-chat-latest
OpenAI ↗gpt-5.2-pro

GPT (Flagship / Thinking) — every version

The full lineage of the GPT (Flagship / Thinking) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.

VersionReleasedContextLicense
GPT-5.5current2026-04-231.05MProprietary
GPT-5.42026-03-05Proprietary
GPT-5.22025-12-11Proprietary
GPT-5.12025-11-12Proprietary
GPT-52025-08-07Proprietary
GPT-4o2024-05-13Proprietary

FAQ

When was GPT-5.2 released?

OpenAI released GPT-5.2 on December 11, 2025, rolling it out first to paid ChatGPT plans and to developers in the API, shortly after a reported internal "code red" to respond to Google's Gemini 3.

What are the differences between GPT-5.2 Instant, Thinking, and Pro?

Instant (gpt-5.2-chat-latest) is the fast everyday model. Thinking (gpt-5.2) is the reasoning model for complex multi-step work, coding, and agents. Pro (gpt-5.2-pro) spends the most inference-time compute for the hardest technical and analytical tasks. All three share a 400K context window and text-plus-vision input.

How much does the GPT-5.2 API cost?

GPT-5.2 Thinking is priced at $1.75 per 1M input tokens, $0.175 per 1M cached input tokens, and $14.00 per 1M output tokens. GPT-5.2 Pro costs more at $21 input and $168 output per 1M tokens.

What is GPT-5.2's context window and knowledge cutoff?

GPT-5.2 has a 400,000-token context window, can output up to 128,000 tokens, and has a knowledge cutoff of August 31, 2025.