Model Drop

o3-pro

June 10, 2025 2 min read

o3-pro, which OpenAI shipped on June 10, 2025, is a version of o3 that thinks longer to give its most reliable answers, replacing o1-pro for ChatGPT Pro users and arriving in the API the same day.

Model: o3-pro (o3-pro)
Maker: OpenAI
Ship date: June 10, 2025
Pricing: $20 / $80 per M tokens
Context: 200K tokens
Weights: Closed
Where you can use it:
- ChatGPT Pro and Team; Enterprise, Edu later
- The API, through the Responses API
When to reach for it:
- Hard problems where reliability beats speed
- Plans built from lots of your own context
- Skip it for quick chats; replies take minutes
How it compares:
- Replaces o1-pro, at 87% less in the API
- Up against Claude Opus 4 and Gemini 2.5 Pro
What makes it different:
- Same model as o3, more compute per answer
- Has search, files, Python and memory
Early reactions: Ben Hylak wrote on Latent Space that it's much smarter than o3, but you only see it when you feed it a lot more context.
Read more: Model release notes: Launching OpenAI o3-pro, o3 and o4-mini System Card

The benchmarks

OpenAI ran every model at the default (medium) ChatGPT thinking time and compared o3-pro only with o3 (medium) and o1-pro; its Codeforces runs for o3 and o3-pro used a newer, harder question set than o1-pro's, the 4/4 reliability rows count a question as solved only if all four tries are right, and no chart shows o3-pro trailing. In the Aider and AA rows, Gemini 2.5 Pro is the 06-05 preview.

OpenAI's numbers

AIME 2024 (competition math): 93%. +3 over o3 (medium) (90%), +7 over o1-pro (86%).
GPQA Diamond (PhD-level science): 84%. +3 over o3 (medium) (81%), +5 over o1-pro (79%).
Codeforces (competitive coding Elo): 2748. +231 over o3 (medium) (2517), +1041 over o1-pro (1707).
AIME 2024 4/4 reliability (right on all four tries): 90%. +10 over o3 (medium) (80%), +10 over o1-pro (80%).
GPQA Diamond 4/4 reliability (right on all four tries): 76%. +9 over o3 (medium) (67%), +2 over o1-pro (74%).
Codeforces 4/4 reliability (worst of four tries, Elo): 2301. +290 over o3 (medium) (2011), +878 over o1-pro (1423).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 21.9. +1.7 over o3 (20.2), +1.3 over Claude Opus 4 (thinking) (20.6), +5.8 over Gemini 2.5 Pro (16.1), +9.5 over o1-pro (12.4).
Aider Polyglot (o3-pro at high, multi-language code edits): 84.9%. +3.6 over o3 (high) (81.3%), +1.8 over Gemini 2.5 Pro (32K thinking) (83.1%), +12.9 over Claude Opus 4 (32K thinking) (72.0%).