OpenAI: GPT-5.4 Nano

Generates natural language responses with optimized inference for low-latency, high-throughput scenarios. Uses a distilled variant of the GPT-5.4 architecture with reduced parameter count and quantization techniques to achieve sub-100ms response times while maintaining semantic coherence. Processes text inputs through a transformer decoder with attention mechanisms, returning streaming or batch completions with configurable temperature and token limits.

Processes images (PNG, JPEG, WebP) as input alongside text prompts and generates descriptive or analytical text responses. Implements vision transformer encoding that converts image pixels into embedding tokens, which are concatenated with text token embeddings and processed through the shared transformer decoder. Supports multiple image inputs per request and handles variable image resolutions through adaptive patching.

Returns model outputs as a stream of tokens via Server-Sent Events (SSE) rather than waiting for full completion, enabling real-time display and early termination. Implements token-by-token streaming with optional backpressure handling, allowing clients to pause or cancel mid-generation. Each streamed token includes logprobs, finish_reason, and usage metadata for fine-grained control and cost tracking.

Processes multiple requests in a single API call with per-request cost tracking and usage attribution. Batches requests are queued and processed asynchronously, returning individual responses with granular token counts (prompt tokens, completion tokens, cached tokens). Implements token-level pricing calculation inline, enabling real-time cost monitoring and budget enforcement per request or user.

Caches prompt tokens across multiple requests, reusing cached embeddings for repeated context (e.g., system prompts, documents, conversation history) to reduce token consumption and latency. Implements a content-addressed cache keyed by prompt hash, with automatic cache invalidation on content changes. Cached tokens are billed at 10% of standard rate, enabling significant cost savings for applications with repeated context.

Enforces model outputs to conform to a provided JSON Schema, guaranteeing valid structured data without post-processing. Uses constrained decoding (token-level masking) to prevent the model from generating tokens that would violate the schema, ensuring 100% schema compliance. Supports nested objects, arrays, enums, and complex type definitions, with optional schema validation before generation.