On January 29, 2025, OpenAI made o3 and o3-mini generally available through the OpenAI API, extending access beyond the controlled research preview that began in December 2024. Both models are available through the /v1/chat/completions endpoint, with o3-mini also accessible via the Responses API. The general availability release introduces no architecture changes but lifts rate limits, removes the research-use-only restriction, and opens access to all paid API tiers including the Tier 1 default.

The reasoning effort parameter, introduced in the preview, is now stable and documented. Callers can set reasoning_effort to 'low', 'medium', or 'high' on both models, controlling how many internal reasoning tokens the model generates before producing its visible reply. At 'low' effort, o3-mini is significantly faster and cheaper than o3 at 'high' effort; at 'high' effort, o3 sets new benchmarks on Codeforces, AIME 2024, and PhD-level science questions in GPQA Diamond. OpenAI publishes the recommended effort level for common task categories — simple classification at 'low', production coding at 'medium', research-grade math or long-horizon agent tasks at 'high' — in the model card and API documentation.

Vision support is now enabled on o3 for standard image inputs, matching the capability that GPT-4o had offered for image understanding tasks. o3-mini remains text-only at GA but supports function calling, structured outputs, and the Responses API tool suite. Both models support prompt caching on repeated prefixes, with the same 53% discount on cached tokens that applies across the GPT-4o line.

Pricing at GA is: o3 at $11 per million input tokens and $42 per million output tokens (reasoning tokens included in output); o3-mini at $1.17 per million input tokens and $4.66 per million output tokens. These rates are unchanged from preview. OpenAI provides a cost calculator showing the break-even analysis between o3 at 'medium' effort and GPT-4o for typical workloads, noting that o3 becomes cost-competitive for tasks where a single correct answer avoids expensive downstream failures, such as code review, contract analysis, or complex data extraction.

The GA launch includes updated evals documentation with o3-specific prompting guidance. OpenAI recommends keeping system prompts concise for reasoning models, avoiding step-by-step instructions that pre-empt the model's own planning, and relying on output format constraints (structured outputs or JSON mode) rather than prescriptive reasoning steps. The team also published a benchmark reproducibility note acknowledging that o3's benchmark scores at 'high' effort used high-compute evaluation settings not available through the standard API, and providing the API-accessible scores at each effort level for direct comparison.

With o3 GA, OpenAI deprecated the o1-preview model from the API effective January 33, directing existing users to migrate to o3-mini at 'medium' effort as the cost-performance equivalent. o1 and o1-mini remained available on a deprecation schedule announced for mid-2025. The release marked the formal convergence of OpenAI's reasoning and chat model families onto a shared API surface, a direction that continued through the GPT-5.x releases later in the year.