Model Drop

o3 & o4-mini

April 16, 2025 2 min read

o3 and o4-mini, which OpenAI shipped on April 16, 2025, are a pair of reasoning models, o3 the most capable and o4-mini the fast, low-cost one, and the first that can use every ChatGPT tool and think with images mid-reasoning.

Model: OpenAI o3 (o3)

Maker: OpenAI

Ship date: April 16, 2025

Pricing: $10 / $40 per M tokens

Context: 200K tokens

Weights: Closed

Where you can use it:

ChatGPT Plus, Pro and Team
The API (Chat Completions and Responses)
Codex CLI, launched the same day

When to reach for it:

Hard, multi-step analysis
Skip it where made-up facts are costly

How it compares:

Replaces o1 in ChatGPT
Rivals Gemini 2.5 Pro and Claude 3.7 Sonnet

What makes it different:

Crops, zooms and rotates images as it thinks
Decides on its own when to call tools

Early reactions: TechCrunch reported it hallucinates more than o1 on OpenAI’s PersonQA test, and Transluce caught it making up actions it took.

Model: OpenAI o4-mini (o4-mini)

Maker: OpenAI

Ship date: April 16, 2025

Pricing: $1.10 / $4.40 per M tokens

Context: 200K tokens

Weights: Closed

Where you can use it:

ChatGPT Plus, Pro and Team; free users via Think
The API (Chat Completions and Responses)

When to reach for it:

High-volume math and code
Fast answers where o3 is overkill

How it compares:

Ahead of o3 on AIME 2025 math
Far behind o3 on web browsing

What makes it different:

About a ninth of o3’s price
Thinks with images, like o3

Early reactions: TechCrunch reported it hallucinates more than o1 on OpenAI’s PersonQA test.

The benchmarks

Every row is o3, with o4-mini as a comparison; OpenAI ran all models at high reasoning effort and compared them only with its own o1, o1-pro, o3-mini and deep research; SWE-bench uses 477 of the 500 tasks, Codeforces scores were run with a terminal, and the BrowseComp scores use Python and browsing. ARC Prize tested at medium effort, and o3 returned only 93 of 100 ARC-AGI-1 tasks.

OpenAI's numbers

AIME 2025 (competition math, no tools): 88.9%. −3.8 under o4-mini (92.7%), +9.7 over o1 (79.2%), +2.4 over o3-mini (86.5%).
GPQA Diamond (PhD-level science, no tools): 83.3%. +1.9 over o4-mini (81.4%), +5.3 over o1 (78.0%), +6.3 over o3-mini (77.0%).
SWE-bench Verified (real GitHub fixes): 69.1%. +1.0 over o4-mini (68.1%), +20.2 over o1 (48.9%), +19.8 over o3-mini (49.3%).
Codeforces (competitive coding Elo): 2706. −13 under o4-mini (2719), +815 over o1 (1891), +633 over o3-mini (2073).
MMMU (college-level visual problems): 82.9%. +1.3 over o4-mini (81.6%), +5.3 over o1 (77.6%).
Humanity's Last Exam (no tools): 20.32%. +6.04 over o4-mini (14.28%), +12.20 over o1-pro (8.12%), +6.92 over o3-mini (13.4%).
BrowseComp (agentic web browsing): 49.7%. +21.4 over o4-mini (28.3%), −1.8 under deep research (51.5%), +47.8 over GPT-4o with browsing (1.9%).
TAU-bench Retail (retail agent tasks): 70.4%. +4.8 over o4-mini (65.6%), −0.4 under o1 (70.8%), +12.8 over o3-mini (57.6%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 20.2. +3.5 over o4-mini (16.7), +2.5 over Claude 3.7 Sonnet (thinking) (17.7), +5.2 over Gemini 2.5 Pro Preview (15.0), +5.0 over o1 (15.2).
ARC-AGI-1 (o3 at medium, abstract visual puzzles): 53%.
ARC-AGI-2 (o3 at medium, abstract reasoning puzzles): 2.9%. +0.6 over o4-mini (medium) (2.3%).