OpenAI announced on 17 December 2024 that its o1 reasoning model is now generally available through the API, alongside new developer tools including Realtime API support and Responses API for structured outputs [source].

The o1 model, previously in preview, is designed for complex reasoning tasks in science, coding, and mathematics. OpenAI states it "thinks before it answers" and performs better on benchmarks requiring multi-step logic. The general release includes function calling and structured outputs, features absent from the preview version.

Developers can now access o1 at $15 per million input tokens and $60 per million output tokens. OpenAI also introduced a new o1-mini model at lower pricing for faster, less complex reasoning tasks.

The Realtime API, previously limited to GPT-4o, now supports GPT-4o mini for speech-to-speech applications. OpenAI claims this reduces costs for voice-based implementations while maintaining low latency.

The Responses API is a new endpoint that enforces structured JSON outputs without requiring developers to specify schemas in prompts. OpenAI describes it as "constrained decoding" that guarantees valid JSON matching a provided schema.

OpenAI also announced Preference Fine-Tuning, allowing developers to train models on comparative preference data rather than single correct answers. The company states this method is suited for tasks where multiple valid responses exist and human preference determines quality.

Rate limits for o1 were increased to 10,000 requests per minute for certain tiers. OpenAI indicated that o1 will eventually replace earlier reasoning models in ChatGPT and the API as capabilities expand.

No independent verification of performance claims was provided in the announcement.