OpenAI: GPT-5.2 Chat

Generates conversational responses with selective internal reasoning using an adaptive compute allocation strategy that routes queries to either fast direct inference or extended chain-of-thought processing based on query complexity heuristics. The model dynamically determines when to invoke deeper reasoning without explicit user control, optimizing for latency while maintaining reasoning quality on complex tasks.

Maintains and processes multi-turn conversation history with automatic context windowing and token-aware truncation, allowing the model to reference previous messages while respecting token limits. Uses a sliding window approach that prioritizes recent messages and system context, with optional explicit conversation state management via the messages array API.

Processes images embedded in chat messages (via URL or base64 encoding) and grounds text generation in visual content, enabling the model to answer questions about images, describe visual scenes, read text from images, and perform visual reasoning tasks. Images are tokenized into visual embeddings and fused with text tokens in the attention mechanism, allowing unified multimodal reasoning.

Enables the model to invoke external functions by generating structured function calls based on a developer-provided schema, with built-in validation against the schema and automatic retry logic for malformed calls. The model receives function definitions as JSON schemas, generates function_call objects with arguments, and receives function results to incorporate into subsequent reasoning steps.

Returns model responses as a stream of text chunks via Server-Sent Events (SSE) rather than waiting for full completion, enabling real-time display of generated text as it's produced. Each chunk includes token usage, finish_reason, and logprobs if requested, allowing client-side token counting and early termination of long responses.

Allows fine-grained control over response randomness via temperature parameter (0.0-2.0), where lower values produce deterministic, focused outputs and higher values increase diversity and creativity. The model uses temperature to scale logits before sampling, affecting both the probability distribution and the sampling strategy (e.g., top-k, top-p) applied during generation.

Provides detailed token usage metrics for each API call, including prompt tokens, completion tokens, and cached tokens (if applicable), enabling cost tracking and optimization. Token counts are returned in the response metadata and can be aggregated across multiple calls to monitor usage patterns and estimate costs based on per-token pricing.

Caches frequently-used prompt segments (system prompts, long documents, code files) to reduce token consumption and latency on subsequent requests with identical context. Uses a content-based hashing mechanism to identify cacheable segments, with cache hits reducing both input token cost (90% discount) and processing latency by reusing pre-computed embeddings.

Allows developers to bias the model's token selection by adjusting logit values for specific tokens before sampling, enabling soft constraints on output (e.g., preferring certain keywords, discouraging specific words, or enforcing output format). Logit biases are applied to the probability distribution without hard constraints, allowing the model to override biases if necessary for coherence.

Enforces JSON-formatted output by constraining the model's token selection to only valid JSON tokens, guaranteeing that responses are valid JSON objects or arrays. The model is instructed to output JSON and the sampling is constrained to prevent invalid JSON syntax, eliminating the need for post-processing or validation.