# o4-mini ![Organization](https://img.shields.io/badge/Org-OpenAI-412991) ![Context](https://img.shields.io/badge/Context-200K%20tokens-1f6feb) ![Pricing](https://img.shields.io/badge/Price-%241.10%20%2F%20%244.40%20per%201M-2da44e) ![License](https://img.shields.io/badge/License-Proprietary-d29922) ![Modality](https://img.shields.io/badge/Modality-Text%20%2B%20Vision-8957e5) ![Released](https://img.shields.io/badge/Released-Apr%202025-555) > **Quick answer:** o4-mini is OpenAI's compact, cost-efficient reasoning model, released April 16, 2025 alongside o3. It scores 74.2% on LiveCodeBench (medium), 95.9% on AI2 Reasoning Challenge, and 74.6% on Arena Hard v2 — delivering near-o3 performance on coding and reasoning tasks at 9× lower cost. Priced at $1.10 input / $4.40 output per 1M tokens with a 200K-token context window. ## At a Glance **Where o4-mini leads** - 95.9% AI2 Reasoning Challenge — top-tier multi-discipline science and reasoning - 74.2% LiveCodeBench (medium effort) — competitive with o3-high at a fraction of the cost - 74.6% Arena Hard v2 — strong general-purpose instruction following - Best reasoning-per-dollar in OpenAI's o-series at $1.10/$4.40 per 1M tokens - Multimodal vision support - Scales well with higher effort settings for harder tasks **Where it lags** - Lower ceiling than o3 on the hardest abstract reasoning tasks (ARC-AGI, frontier math) - 200K context window — constrained vs 1M+ long-context models - Closed API model with no open weights - CyberGym score of 2.5% — not suited for cybersecurity tasks **Best for:** Cost-sensitive production pipelines, high-throughput coding assistance, API-integrated agents, and workflows where o3's premium is not justified by task difficulty. ## What o4-mini Is o4-mini is the small, efficient tier of OpenAI's April 2025 o-series release. Like o3, it uses internal chain-of-thought reasoning before producing a final response, with low/medium/high effort settings. The key trade-off vs o3: o4-mini's capability ceiling is lower on the hardest abstract reasoning tasks, but on the coding and science benchmarks that dominate most production agent workloads, the gap is small. The model's 95.9% AI2 Reasoning Challenge score — a multi-discipline science benchmark covering physics, chemistry, and biology — indicates that o4-mini's reasoning capability is genuinely strong, not merely fast. Its LiveCodeBench score of 74.2% at medium effort is within 1.6 points of o3-high, meaning for most coding tasks the cost premium for o3 is rarely justified. For teams building production AI agents, o4-mini is typically the right default in the o-series: capable enough for the majority of real-world reasoning tasks, priced for at-scale operation, and responsive enough at medium effort to meet latency requirements. ## Specifications | Field | Value | |---|---| | Organization | OpenAI | | Parameters | Undisclosed | | Context window | 200,000 tokens | | Max output | 100,000 tokens | | Architecture | Chain-of-thought reasoning (o-series) | | License | Proprietary (API only) | | Release date | April 16, 2025 | | Knowledge cutoff | June 2024 (estimated) | | Modality | Text + Vision (multimodal) | ## Pricing | | Input (per 1M tokens) | Output (per 1M tokens) | |---|---|---| | OpenAI API | $1.10 | $4.40 | Reasoning tokens are billed as output tokens. Cached input receives a 75% discount. Pricing per [OpenAI pricing page](https://openai.com/api/pricing/). ## Context Window o4-mini has a **200,000-token context window** — roughly **150 pages** of text in a single request, matching o3. This is sufficient for most agentic workflows but constrains whole-codebase or long-document tasks where Gemini 2.5 Pro's 1M context is preferable. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [AI2 Reasoning Challenge](/benchmarks/ai2-reasoning-challenge) | 95.9% | Benchgen evaluation | 2025-07 | | [Arena Hard v2](/benchmarks/arena-hard-v2) | 74.6% | Benchgen evaluation | 2025-07 | | [LiveCodeBench](/benchmarks/livecodebench) | 74.2% (medium) | Benchgen evaluation | 2025-07 | | [Humanity's Last Exam](/benchmarks/humanitys-last-exam) | 14.28% | Benchgen evaluation | 2025-07 | | [CyberGym](/benchmarks/cybergym) | 2.5% | Benchgen evaluation | 2025-07 | | [SHADE-Arena](/benchmarks/shade-arena) | 6.5 overall success | [Anthropic research post](https://www.anthropic.com/research/shade-arena-sabotage-monitoring) | 2025-06 | *Scores from Benchgen evaluations or attributed third-party sources. Figures depend on effort setting and harness — see the source for methodology.* ## o4-mini vs Alternatives | Model | Context | LiveCodeBench | AI2 RC | Price (in/out per 1M) | |---|---|---|---|---| | [o3](/models/openai/o3) | 200K | 75.8% (high) | — | $10 / $40 | | **o4-mini** | 200K | 74.2% (medium) | 95.9% | $1.10 / $4.40 | | [GPT-5.6 Luna](/models/openai/gpt-5-6-luna) | 1.1M | — | — | $1 / $6 | | [Gemini 2.5 Pro](/models/google/gemini-2-5-pro) | 1M | 73.6% | — | $1.25 / $10 | | [DeepSeek R1](/models/deepseek/r1) | 128K | — | 96.4% | ~$0.55 / $2.19 | o4-mini's strongest case: it matches o3-high on LiveCodeBench at 9× lower cost. vs Gemini 2.5 Pro, o4-mini is cheaper and slightly stronger on coding tasks, but Gemini offers a 5× larger context window. ## How o4-mini Performs on Real Agent Tasks o4-mini's 95.9% AI2 Reasoning Challenge and 74.2% LiveCodeBench position it as the default reasoning model for production workloads requiring science and coding capability. Its CyberGym score of 2.5% indicates it is not suited for cybersecurity tasks — for those, GPT-5.6 Sol is the more appropriate choice. For most teams, o4-mini at medium effort is the right starting point. Upgrade to high effort when evaluation data shows a consistent accuracy gap; switch to o3 only when task difficulty consistently pushes past o4-mini-high's ceiling. ## Use o4-mini via API ```python from openai import OpenAI client = OpenAI(api_key="YOUR_API_KEY") response = client.chat.completions.create( model="o4-mini", messages=[{"role": "user", "content": "Debug this Python function and explain the fix..."}], reasoning_effort="medium", ) print(response.choices[0].message.content) ``` ## Related Models & Benchmarks - Compare: [o3](/models/openai/o3) · [GPT-5.6 Luna](/models/openai/gpt-5-6-luna) · [Gemini 2.5 Pro](/models/google/gemini-2-5-pro) · [DeepSeek R1](/models/deepseek/r1) - Benchmarks: [LiveCodeBench](/benchmarks/livecodebench) · [AI2 Reasoning Challenge](/benchmarks/ai2-reasoning-challenge) · [Arena Hard v2](/benchmarks/arena-hard-v2) · [Humanity's Last Exam](/benchmarks/humanitys-last-exam) ## Frequently Asked Questions

What is o4-mini?

o4-mini is OpenAI's compact chain-of-thought reasoning model, released April 16, 2025. It uses internal reasoning before responding — with low/medium/high effort settings — delivering near-o3 performance on coding and science tasks at 9× lower cost.

What is o4-mini's context window?

o4-mini supports a 200,000-token context window — roughly 150 pages of text — with up to 100,000 output tokens.

How much does o4-mini cost?

o4-mini is priced at $1.10 per 1M input tokens and $4.40 per 1M output tokens. Reasoning tokens are billed as output. Cached inputs receive a 75% discount.

Is o4-mini open source?

No. o4-mini is a proprietary, API-only model. There are no open weights.

What is the difference between o4-mini and o3?

o3 ($10/$40 per 1M) has a higher capability ceiling on the hardest reasoning tasks. o4-mini ($1.10/$4.40 per 1M) matches o3 on most coding and science benchmarks at 9× lower cost. For most production workloads, o4-mini at medium or high effort is the correct default.

What is o4-mini's knowledge cutoff?

o4-mini's training knowledge cutoff is estimated at June 2024.

--- *Specs and scores sourced from OpenAI's official o4-mini announcement (April 16, 2025) and Benchgen evaluations; third-party benchmark scores attributed inline. Pricing cited to the OpenAI pricing page. Last updated 2026-07-23.*

o4-mini — Evaluations

Evaluations