o3

Use the model ID o3 through the compatible chat completions endpoint. Check the current API documentation for the exact endpoint path and any required headers.

To get a cache hit on o3, send requests with identical prompt prefixes. The cache is keyed by the beginning of your prompt; as long as the prefix matches a previous request, subsequent calls benefit from the lower cached rate. Keep system messages and initial context consistent across calls for best results.

Test your most demanding multi-step reasoning tasks with text, code, and image inputs. Confirm the model’s accuracy and latency meet your requirements. Run the same integration tests used for your current client.

For o3, set the max_tokens or max_completion_tokens parameter to control spending per call. Start with a budget aligned to your typical response length and increase only when longer outputs are genuinely needed. Monitor token usage in the API response headers to tune your budget over time.

o3 supports prompt caching, which can significantly reduce costs for repeated prompts with shared prefixes. The cache read price is substantially lower than the standard input price. Check the current pricing page for the exact cache hit discount ratio for o3.

Compare o3 against models of similar capability and pricing on your actual workload. Evaluate latency, output quality, instruction following, and cost per task. Run side-by-side tests with your own prompts — the best model depends on your specific use case, not just benchmark scores.