Model Drop

GPT-5.1

November 12, 2025 1 min read

GPT-5.1, which OpenAI shipped on November 12, 2025, is an upgrade to GPT-5 that adapts how long it thinks to each question and answers in a warmer default tone.

Model: GPT-5.1 (gpt-5.1)
Maker: OpenAI
Ship date: November 12, 2025
Pricing: $1.25 / $10 per M tokens
Context: 400K tokens
Where you can use it:
- ChatGPT, paid plans first, then free
- OpenAI API from November 13
When to reach for it:
- Agentic coding at GPT-5 prices
- Fast replies with reasoning turned off
- Skip it if you want terse, neutral replies
How it compares:
- Edges Grok 4 and Claude Sonnet 4.5 overall
- Well ahead of Gemini 2.5 Pro
What makes it different:
- Thinks less on easy asks, more on hard ones
- New "none" reasoning mode in the API
Early reactions: Reviewers saw an incremental upgrade with better instruction following, and were split on the warmer, chattier default tone.
Read more: GPT-5.1: A smarter, more conversational ChatGPT, Introducing GPT-5.1 for developers, GPT-5.1 system card addendum

The benchmarks

OpenAI ran GPT-5.1 and GPT-5 both at high reasoning effort and compared GPT-5.1 only against GPT-5; on τ²-bench Telecom it gave GPT-5.1 a short, generically helpful prompt.

OpenAI's numbers

SWE-bench Verified (real GitHub fixes): 76.3%. +3.5 over GPT-5 (high) (72.8%).
GPQA Diamond (grad-level science): 88.1%. +2.4 over GPT-5 (high) (85.7%).
AIME 2025 (competition math): 94.0%. −0.6 under GPT-5 (high) (94.6%).
FrontierMath (expert math, with Python): 26.7%. +0.4 over GPT-5 (high) (26.3%).
MMMU (visual reasoning): 85.4%. +1.2 over GPT-5 (high) (84.2%).
τ²-bench Retail (customer-service agent): 77.9%. −3.2 under GPT-5 (high) (81.1%).
τ²-bench Telecom (customer-service agent): 95.6%. −1.1 under GPT-5 (high) (96.7%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 24.7. +2.2 over Grok 4 (22.5), +4.0 over Claude Sonnet 4.5 (20.7), +8.6 over Gemini 2.5 Pro (16.1).
ARC-AGI-2 (ARC Prize semi-private, high effort): 17.6%. +1.6 over Grok 4 (16.0%), +4.0 over Claude Sonnet 4.5 (13.6%), +7.7 over GPT-5 (high) (9.9%).