HomeAI models
OpenAI: o4 Mini High
OpenAI
•
o4-mini (High)
•
Released 
April 2025

OpenAI
o4-mini (High)

Small OpenAI reasoning model at high effort: strong math and coding with image input, retiring October 23, 2026.

Modality:
Text
Image
PDF
model ID
openai/o4-mini-high
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

N/A
tok/s

Intelligence Index *

16.7
/ 100

Context Window *

200000
tokens

Input price

6.6
Anytoken

Output price

26.4
Anytoken
o4-mini (High): Maximum-Effort Reasoning in a Small, Fast Model o4-mini (High) is OpenAI's small o-series reasoning model run with reasoning_effort set to high. It is the same underlying model as o4-mini. OpenAI built it for math, coding, and visual tasks, and it sits below o3 in the lineup. It accepts text and image input and has a 200K context window. Its best fit is asynchronous reasoning on hard problems. OpenAI has scheduled the API shutdown for October 23, 2026, so it suits existing workloads and migration benchmarking more than new builds. Test o4-mini (High) on AnyAPI.ai

Performance

Math and Coding Strength Behind a Long Reasoning Delay

OpenAI reports o4-mini reaches 99.5% pass@1 on AIME 2025 with a Python interpreter, and 68.1% on SWE-bench Verified. These are provider-reported figures, but they show why the model was positioned for math, coding and visual reasoning at small-model scale. The high setting spends more hidden reasoning tokens per request. That tends to help on hard problems but adds delay and uses the output budget. In production, use it for batch or asynchronous reasoning, not interactive replies.

Benchmarks

Independent Measurements: Fast Output, Slow First Token

Artificial Analysis measured o4-mini (high) at 146.5 tokens per second, but with a 30.79-second time to first token versus 1.05 seconds for GPT-4.1. On the same comparison page it scores 17% on Humanity's Last Exam versus 4% for GPT-4.1, and 61% on AA-LCR long-context reasoning versus 68%. Speed and latency vary by test date, and Artificial Analysis labels some o4-mini (high) figures as estimated.

Output Speed

*
N/A
tok/s

Intelligence Index

*
16.7
/ 100

MMLU *

Broad world knowledge and problem-solving
83
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
78
%

HLE *

Adherence to multi-step structured instructions.
17
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
86
%

Technical Specifications

What the model supports

o4-mini (High) takes text and image input and returns text only. Audio and video are not supported. The 200,000-token context window and 100,000-token maximum output are separate limits, and reasoning tokens count toward the output limit. At high effort, hard prompts can spend much of that allowance on hidden reasoning, so set generous output limits. The most important production fact is lifecycle: OpenAI lists a shutdown date of October 23, 2026.
Verified Specifications — 
o4-mini (High)
*
Input modalities
Text
Image
PDF
output modalities
Text
Context window
200000
 tokens
Maximum output tokens
100000
Reasoning
Yes
Knowledge cutoff
April 2025
Pricing (standard)
6.6
 AnyTokens in
 / 
26.4
 AnyTokens out

Comparison

o4-mini (High) or GPT-5 mini: Is the Migration Worth It?

OpenAI's model page describes o4-mini as succeeded by GPT-5 Mini, and both are small, cost-oriented reasoning models. On Artificial Analysis's comparison, both score 17 on the Intelligence Index (estimated for o4-mini), so quality is close on that measure. The difference is speed: 142.2 tokens per second for o4-mini (high) versus 97.9 for GPT-5 mini (high). GPT-5 mini is also scheduled for shutdown on December 11, 2026, so neither is a long-term target.

Dimension
o4-mini (High)
GPT-5 mini (High)
Context window *
200000
tokens
400000
tokens
Output speed *
N/A
tok/s
0.00
tok/s
Intelligence Index *
16.7
16.8
Input pricing
6.6
AnyToken
1.5
AnyToken
Output pricing
26.4
AnyToken
12
AnyToken
Knowledge cutoff *
April 2025
August 2025

Choose o4-mini (High) when you are keeping an existing, evaluated pipeline running until the shutdown and want higher measured output speed on long reasoning outputs. Choose GPT-5 mini when you want the model OpenAI names as the successor and a longer runway before retirement. For anything planned beyond 2026, evaluate OpenAI's listed substitute for o4-mini, gpt-5.6-terra, instead of either model.

Limitations & Trade-offs

Where o4-mini (High) falls short

1
Retirement on October 23, 2026. OpenAI's deprecations page lists o4-mini and the o4-mini-2025-04-16 snapshot for shutdown on October 23, 2026, and after that date requests to these IDs fail. New products should not start here. Existing pipelines should be benchmarked against the listed substitute now, and tested for parameter differences: one developer report found function tools combined with reasoning effort rejected on Chat Completions for that substitute.
2
Long time to first token. Artificial Analysis measured 30.79 seconds to first token and 34.20 seconds end-to-end for o4-mini (high), versus 1.05 and 4.17 seconds for GPT-4.1. The model reasons before emitting visible tokens, and streaming does not hide that delay. Chat interfaces, autocomplete and voice agents are poor fits. Prefer a non-reasoning model for interactive latency, or lower the reasoning effort when full depth is unnecessary.
3
Reasoning tokens consume the output budget. The 100,000-token output cap includes reasoning tokens, and reasoning tokens are billed at the same rate as output tokens. High effort therefore makes output length and cost less predictable than with non-reasoning models. Budget max output tokens generously, monitor reasoning-token usage per request, and use lower effort for routine tasks such as classification or short extraction.
4
Text-only output and limited long-context depth. The model accepts text and images but produces text only, with no audio or video support. On Artificial Analysis's AA-LCR long-context test it scored 61% versus 68% for GPT-4.1, which has a 1,000k-token window against 200k. Whole-codebase analysis or very long document retrieval favors a longer-context model.

Best-Fit Workloads

Where this model earns its place

01

Competition-style math and quantitative reasoning

‍
OpenAI positioned o4-mini as its best-performing benchmarked model on AIME 2024 and 2025 at launch, with 99.5% pass@1 on AIME 2025 when given a Python interpreter. That supports asynchronous solvers, verification pipelines and quantitative analysis where high effort pays off. Tool-assisted scores are not comparable to no-tool scores, and the model's retirement makes it better as a baseline than a new dependency.

02

Batch debugging and patch generation

‍
OpenAI reports 68.1% on SWE-bench Verified at high reasoning effort on a 477-task subset, and function calling is supported for agent loops. This suits CI-time code repair, review bots and issue triage that can tolerate a delay before the first token. Interactive IDE completion is a poor fit given the measured 30-second first-token latency. Treat vendor SWE-bench scores as scaffold-dependent.

03

Chart, diagram and screenshot reasoning

‍
OpenAI describes o4-mini as strong in visual tasks, and the API accepts image input alongside text. That fits analysis of charts, technical diagrams or UI screenshots followed by a reasoned text answer, such as financial-report review or bug reproduction from screenshots. Output is text only, and video and audio inputs are not supported.

04

Migration baselines for retiring pipelines

‍
Teams with prompts already tuned for o4-mini can use o4-mini (High) as a reference when evaluating OpenAI's listed substitute. Replaying identical prompts lets you compare accuracy, latency, structured-output validity and tool-call behavior before the October 23, 2026 shutdown. This is a transitional use, not a recommendation for new production architecture.

Pricing in anytokens via AnyAPI
Input
6.6
₳
Output
26.4
₳
Cache write
—
₳
Cache read
1.65
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsComparisonLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access o4-mini (High) via AnyAPI.ai

Access o4-mini (High) through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate o4-mini (High) without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access o4-mini (High) and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test o4-mini (High) against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use o4-mini (High) from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use o4-mini (High) for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

o4-mini (High) has a 200,000-token context window and a 100,000-token maximum output. Reasoning tokens count toward the output limit. Input can be text or images, and output is text only.

Yes. OpenAI's deprecations page lists o4-mini and o4-mini-2025-04-16 for shutdown on October 23, 2026, with gpt-5.6-terra as the recommended substitute. After that date, requests to these model IDs fail, so retest prompts, tool calls and structured outputs on the replacement beforehand.

o4-mini (High) is the same model as o4-mini with reasoning_effort set to high. The API exposes low, medium and high effort levels. High spends more reasoning tokens, which can improve accuracy on hard problems but increases latency and output token usage.

Artificial Analysis measured about 146.5 output tokens per second for o4-mini (high), but a 30.79-second time to first token. It streams quickly once it begins but delays the first visible token while reasoning, so it suits batch or asynchronous jobs better than real-time chat. Results vary by test date.

Yes. OpenAI documents text and image input with text output, plus streaming, function calling and structured outputs. Audio and video are not supported, and predicted outputs are not supported.

* Benchmark data source: Artificial Analysis artificialanalysis.ai