o4-mini

o4-mini belongs to OpenAI's o-series of reasoning models, engineered to think through problems more thoroughly before responding. The model stands apart from standard language models through its agentic design—capable of autonomously deploying multiple tools in combination, including web search, Python-powered data analysis, visual input reasoning, and image generation. Built to understand when and how to leverage each capability, it produces detailed, multi-step solutions to complex problems, typically completing this extended reasoning process in under a minute. This architectural approach targets users ranging from curious individuals exploring AI to researchers tackling advanced technical challenges.

Third-party evaluation from METR, conducted prior to the model's release, found that o4-mini achieved 50% time-horizon performance approximately 1.5 times that of competing reasoning systems on the HCAST autonomy benchmark—a metric measuring how far into complex, multi-step tasks the model can reliably progress. The model also demonstrated strong results on RE-Bench, a benchmark designed to assess AI capabilities in software development and R&D workflows. These improvements reflect refinements in training that better calibrate the model's reasoning chains and tool-use decisions, positioning it as a practical choice for applications requiring both depth of analysis and the flexibility to orchestrate multiple capabilities together.

OpenAIo4-minio-minideprecated

Open provider page Provider docs

Quick Info

Provider: OpenAI
Model key: o4-mini
Release date: Apr 16, 2025
Last updated: Apr 16, 2025
Knowledge cutoff: 2024-05
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $1.10
Output token cost: $4.40

Limits

Output tokens: 100,000 tokens
Context window: 200,000 tokens

OpenCode

Model variants

Fast

Input

$2.00

per 1M tokens

Output

$8.00

per 1M tokens

Cache Read

$0.50

per 1M tokens

Flex

Input

$0.55

per 1M tokens

Output

$2.20

per 1M tokens

Cache Read

$0.138

per 1M tokens

Transparent token rates

Compare o4-mini pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about o4-mini

DevPass (LLM Gateway)

CoverageBenchmark

Independent benchmark aggregator llm-stats.com (page date September 22, 2026) ranks OpenAI's o4-mini at #146 on its composite LLM Stats Score, with category standings including Chat (58 of 133), Healthcare (70 of 245), Vision (91 of 210), Math (97 of 327), Reasoning (138 of 366), Writing (62 of 114), Search (65 of 90), The page enumerates specific benchmark scores for o4-mini sourced to openai.com, including AIME 2024 (rank 2, score 0.93/1, accuracy without tools), AIME 2025 (rank 36, score 0.93/1, accuracy without tools), MathVista (rank 5, score 0.84/1), and MMMU. These figures provide a quantitative snapshot of o4-mini's mathemati

OpenRouter

Coverage

Wikiwand's encyclopedia entry for OpenAI o4-mini corroborates the model's timeline: initial release on April 16, 2025, broader rollout including free-tier ChatGPT users on April 24, 2025, retirement announced January 29, 2026, and effective ChatGPT retirement on February 13, 2026. It identifies the predecessor as o3-mi The entry confirms o4-mini was a vision-language model capable of processing text and images, including analyzing whiteboard sketches during its chain-of-thought phase. Both o4-mini and o4-mini-high were released as vision-language models, distinguishing o4-mini from earlier text-only o-series entries. The page is a th

Model Oracle AI

Coverage

A Hacker News comment dated April 16, 2025 from an OpenAI employee (tedsanders) clarifies that in ChatGPT, o4-mini is replacing o3-mini as a straight 1-to-1 upgrade, while in the API o4-mini is introduced as a new model option with o3-mini continuing to be supported so developers who built products atop o3-mini can mai The same thread explains the rationale behind keeping both models available: launching a new model would otherwise risk breaking production apps, and developers are given the option to test both and switch when they like. The commenter also notes that naming and offering improvements were being considered later in 2025