Model Details
o4-mini (high)
Quick answer: o4-mini at high reasoning effort scores 80.2% on LiveCodeBench, 18.08% on Humanity's Last Exam, and 35.5% on AetherCode — outperforming o3 (88%, 20.32%) on coding while costing 9× less. At $1.10/$4.40 per 1M tokens with vision support, it offers the best reasoning benchmark performance per dollar of any model tracked on Benchgen at its launch.
At a Glance
Where o4-mini (high) leads
- 80.2% LiveCodeBench — exceeds o3's 75.8% at 9× lower cost
- 35.5% AetherCode — strongest agentic coding score in the mini price tier
- 18.08% Humanity's Last Exam — near o3's 20.32% at a fraction of the price
- Vision support — multimodal reasoning at mini-tier pricing
- $1.10/$4.40 per 1M tokens — extraordinary price/performance
Where it lags
- Lower HLE than o3 (20.32%) and Gemini 2.5 Pro (21.64%)
- High effort = more reasoning tokens = longer latency
- Not suitable for low-latency, high-throughput pipelines at high effort
- No open weights
Best for: Cost-efficient competitive programming, technical problem-solving, agentic coding tasks, and any pipeline where o3-tier accuracy is needed at o4-mini pricing.
What o4-mini (high) Is
o4-mini at high effort is the maximum-accuracy configuration of OpenAI's April 2025 compact reasoning model. "High" allocates the largest internal reasoning token budget, trading latency for accuracy. The result is LiveCodeBench 80.2% — higher than o3's 75.8% at a price 9× lower.
The o4-mini family brought two major advances over o3-mini: improved base reasoning capability, and multimodal vision support. At high effort, these translate into benchmark results that challenge the previous generation's flagship model (o3) on coding tasks.
For agentic coding workloads, o4-mini-high's 35.5% AetherCode score is the standout result. AetherCode measures end-to-end agentic code generation across real-world software engineering tasks — 35.5% represents strong performance for a model at this price point.
Specifications
| Field | Value |
|---|---|
| Organization | OpenAI |
| Parameters | Undisclosed |
| Context window | 200,000 tokens |
| Effort setting | High (maximum reasoning budget) |
| 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 billed as output. At high effort, reasoning token counts are higher, increasing effective cost. Pricing per OpenAI pricing page.
Context Window
o4-mini (high) has a 200,000-token context window — suitable for long document analysis and multi-file code review tasks.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| LiveCodeBench | 80.2% (high effort) | Benchgen evaluation | 2025-07 |
| Humanity's Last Exam | 18.08% | Benchgen evaluation | 2025-07 |
| AetherCode | 35.5% | Benchgen evaluation | 2025-07 |
o4-mini (high) vs Alternatives
| Model | LiveCodeBench | HLE | AetherCode | Price (in/out per 1M) |
|---|---|---|---|---|
| o4-mini (high) | 80.2% | 18.08% | 35.5% | $1.10 / $4.40 |
| o3 | 75.8% | 20.32% | — | $10 / $40 |
| o4-mini (med) | 74.2% | 14.28% | — | $1.10 / $4.40 |
| Gemini 2.5 Pro | 73.6% | 21.64% | 32.7% | $1.25 / $10 |
| o3-mini (high) | 67.4% | — | — | $1.10 / $4.40 |
o4-mini (high) vs o3: higher LiveCodeBench and AetherCode at 9× lower cost — o3 retains a slight HLE edge (20.32% vs 18.08%). For coding-heavy workloads, o4-mini-high is the clear winner. For general expert reasoning, o3 or Gemini 2.5 Pro may be preferable.
Use o4-mini (high) via API
from openai import OpenAI
client = OpenAI(api_key="YOUR_API_KEY")
response = client.chat.completions.create(
model="o4-mini",
reasoning_effort="high",
messages=[{"role": "user", "content": "Implement a Dijkstra solver with priority queue..."}],
)
print(response.choices[0].message.content)Related Models & Benchmarks
- Compare: o4-mini · o3 · o3-mini (high) · Gemini 2.5 Pro
- Benchmarks: LiveCodeBench · Humanity's Last Exam · AetherCode
Frequently Asked Questions
What is o4-mini (high)?
o4-mini at high effort is OpenAI's April 2025 compact reasoning model at maximum accuracy, scoring 80.2% LiveCodeBench — higher than o3 (75.8%) at 9× lower cost.What is the difference between o4-mini and o4-mini high?
The effort setting controls how many internal reasoning tokens the model uses. o4-mini (high) scores 80.2% LiveCodeBench vs 74.2% for the medium default, at higher latency and slightly higher effective token cost. Choose high effort for accuracy-critical tasks, medium for latency-sensitive workloads.How much does o4-mini (high) cost?
The base rate is $1.10 input / $4.40 output per 1M tokens. At high effort, the model generates more reasoning tokens, increasing the effective output cost per request.Is o4-mini (high) better than o3?
For coding tasks (LiveCodeBench 80.2% vs 75.8%) and AetherCode (35.5%), yes — at 9× lower cost. For general expert reasoning (HLE), o3 retains a slight edge (20.32% vs 18.08%).Specs and scores sourced from OpenAI's official o4-mini announcement (April 2025) and Benchgen evaluations. Pricing cited to the OpenAI pricing page. Last updated 2026-07-24.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.