Model Drop

GPT-5.2

December 11, 2025 1 min read

GPT-5.2, which OpenAI shipped on December 11, 2025, is OpenAI's flagship for professional knowledge work such as spreadsheets, slide decks, code and long documents.

Model: GPT-5.2 (gpt-5.2)
Maker: OpenAI
Ship date: December 11, 2025
Pricing: $1.75 / $14 per M tokens
Context: 400K tokens
Where you can use it:
- ChatGPT paid plans: Instant, Thinking, Pro
- OpenAI API, plus GPT-5.2 Pro
When to reach for it:
- Spreadsheets, decks and long reports
- Hard problems at the top effort setting
- Skip it for quick, casual questions
How it compares:
- Narrowly ahead of Claude Opus 4.5 overall
- Far ahead of Gemini 3 Pro on puzzles
What makes it different:
- New "xhigh" reasoning effort level
- Holds up across very long documents
Early reactions: Testers praised it on hard, long-running problems but found Thinking mode slow and wordy for everyday questions.
Read more: Introducing GPT-5.2, GPT-5.2 system card update

The benchmarks

OpenAI ran GPT-5.2 Thinking at xhigh effort (heavy for GDPval) against GPT-5.1 Thinking at high, compared only against its own models, and listed GPT-5 as the GDPval baseline.

OpenAI's numbers

GDPval (expert knowledge work, wins or ties): 70.9%. +32.1 over GPT-5 (38.8%).
SWE-bench Pro (harder multi-language fixes): 55.6%. +4.8 over GPT-5.1 Thinking (50.8%).
SWE-bench Verified (real GitHub fixes): 80.0%. +3.7 over GPT-5.1 Thinking (76.3%).
GPQA Diamond (grad-level science): 92.4%. +4.3 over GPT-5.1 Thinking (88.1%), −0.8 under GPT-5.2 Pro (93.2%).
ARC-AGI-2 (abstract puzzles): 52.9%. +35.3 over GPT-5.1 Thinking (17.6%), −1.3 under GPT-5.2 Pro (high) (54.2%).
Humanity's Last Exam (no tools) (expert questions): 34.5%. +8.8 over GPT-5.1 Thinking (25.7%).
BrowseComp (hard web research): 65.8%. +15.0 over GPT-5.1 Thinking (50.8%), −12.1 under GPT-5.2 Pro (77.9%).
τ²-bench Telecom (customer-service agent): 98.7%. +3.1 over GPT-5.1 Thinking (95.6%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 30.4. +1.3 over Claude Opus 4.5 (29.1), +2.4 over Gemini 3 Pro (28.0).
ARC-AGI-2 (ARC Prize semi-private, xhigh effort): 52.9%. +7.8 over Gemini 3 Deep Think (45.1%), +15.3 over Claude Opus 4.5 (37.6%), +21.8 over Gemini 3 Pro (31.1%).