OpenAI: o3 Pro

Implements reinforcement learning-trained reasoning that allocates variable computational budget across thinking phases before generating responses. The model uses an internal chain-of-thought mechanism where it can 'think' for extended periods (up to specified token limits) before committing to an answer, similar to o1/o3 architecture. This enables structured problem decomposition, hypothesis testing, and self-correction within a single inference pass without requiring external orchestration.

Accepts both text and image inputs in a single API call, processing visual content through a vision encoder that extracts semantic features before feeding them into the reasoning pipeline. The model can analyze images, diagrams, charts, and screenshots, then apply its extended reasoning capabilities to answer questions about visual content or solve problems that combine textual and visual information.

Supports JSON schema-based output constraints that force the model to generate responses conforming to a specified structure. The model's reasoning process is aware of the output schema, allowing it to plan solutions that fit the required format before generating. This enables reliable extraction of structured data, function arguments, or domain-specific formats without post-processing or retry logic.

Maintains conversation history across multiple turns, with each turn's reasoning and output contributing to the model's understanding of subsequent queries. The model can reference previous reasoning steps, correct earlier conclusions, and build on prior analysis without requiring explicit context injection. Thinking tokens are computed per-turn, allowing the model to allocate reasoning budget based on conversation state.

Generates code solutions by reasoning through algorithmic approaches, edge cases, and implementation details before producing output. The model can analyze existing code, identify bugs, suggest optimizations, and generate complete implementations for complex algorithms. Reasoning is applied to understand problem constraints and design decisions before code is written, reducing hallucinations and improving correctness.

Solves mathematical problems by reasoning through problem decomposition, intermediate calculations, and solution verification. The model can handle algebra, calculus, number theory, combinatorics, and applied mathematics by explicitly working through each step. Reasoning allows the model to catch calculation errors and verify solutions before output, improving accuracy on complex multi-step problems.

Provides confidence assessments and uncertainty estimates alongside reasoning outputs, allowing the model to explicitly acknowledge when it is less certain about conclusions. The reasoning phase includes exploration of alternative interpretations and confidence in different solution paths, which can be surfaced to the user. This enables better decision-making when the model's output will be used in high-stakes contexts.

Exposes o3-pro through OpenAI's REST API with detailed token accounting that separates thinking tokens from output tokens. Clients can track usage in real-time, estimate costs before making requests, and optimize spending by adjusting thinking budget. The API returns detailed metadata about token consumption, allowing builders to understand the cost-benefit trade-off of extended reasoning.