OpenAI o3 - chunhualiao/public-docs GitHub Wiki
Here's a detailed comparison of OpenAI's o3 and o3-mini models, highlighting their capabilities, pricing, and access specifications:
| Feature | o3 | o3-mini |
|---|---|---|
| Release Date | April 16, 2025 | January 31, 2025 |
| Model Type | Flagship reasoning model with advanced capabilities in coding, math, science, and vision | Cost-efficient reasoning model optimized for STEM tasks; does not support vision |
| Context Window | Up to 200,000 tokens | Up to 200,000 tokens |
| Output Limit | Up to 100,000 tokens per response | Up to 100,000 tokens per response |
| Pricing (per 1M tokens) | Input: $10.00Cached Input: $2.50Output: $40.00 | Input: $1.10Cached Input: $0.275Output: $4.40 |
| Reasoning Effort Levels | Fixed reasoning effort | Selectable levels: Low, Medium, High |
| Supported Features | Web browsingImage generation and visual understandingFunction callingStructured outputsStreaming responsesDeveloper messages | Function callingStructured outputsStreaming responsesDeveloper messagesSearch integration for web-sourced answers |
| API Access | Available via Responses API | Available via Chat Completions, Assistants, and Batch APIs |
| Access Requirements | API usage tiers 4 and 5Tiers 1–3 can gain access through API organization verification | API usage tiers 1 through 5Available to ChatGPT Plus, Team, and Pro usersLimited access for free ChatGPT users |
| Performance Benchmarks | GPQA Diamond: 87.7% accuracyCodeforces Elo Rating: 2727ARC-AGI Benchmark: 87.5% accuracy | AIME 2024 (High Effort): 83.6% accuracyGPQA Diamond (High Effort): 77.0% accuracyCodeforces Elo Rating (High Effort): 2073SWE-bench Verified (High Effort): 48.9% accuracy |
| Safety Features | Incorporates deliberative alignment for enhanced safety and adherence to human-written safety specifications | Utilizes deliberative alignment; demonstrates significant improvements over previous models in safety and jailbreak evaluations |
| Latency | Higher latency due to complex reasoning capabilities | Approximately 24% faster response time compared to o1-mini; average response time of 7.7 seconds with medium reasoning effort |
| Use Cases | Ideal for tasks requiring deep reasoning, complex problem-solving, and multimodal understanding, including visual tasks | Best suited for STEM-related tasks such as math, coding, and science, where cost-efficiency and speed are prioritized over multimodal capabilities |
Notes:
-
Cached Input Tokens:These are repeated inputs that the model has previously processed, allowing for reduced costs
-
Batch API Discount:Utilizing the Batch API can provide up to a 50% discount on input and output token costs
-
Access Tiers:Access to these models may depend on your API usage tier. o3 is available to users in tiers 4 and 5, while o3-mini is accessible to tiers 1 through 5
For more detailed information and to access these models, visit the [OpenAI API Pricing page](https://openai.com/api/pricing/).