OpenAI: GPT-5.4 Mini

Processes both natural language text and image inputs through a shared transformer architecture that encodes visual and textual information into a unified representation space. The model uses vision transformer (ViT) patches for image tokenization and merges them with text tokens in a single attention mechanism, enabling cross-modal reasoning where image context directly influences text generation and vice versa.

Implements structured reasoning through intermediate thinking steps that are computed efficiently within the model's forward pass, using a sparse attention pattern that prioritizes reasoning tokens over raw output. The model learns to decompose complex problems into logical sub-steps, with each step building on previous reasoning without requiring separate API calls or external orchestration.

Generates and analyzes code across 40+ programming languages by internally representing code as abstract syntax trees (ASTs) rather than raw text tokens. The model understands structural relationships between code elements (function definitions, control flow, variable scope) and can perform refactoring, bug detection, and cross-language transpilation by reasoning about AST transformations rather than pattern matching on syntax.

Enables the model to invoke external functions and APIs by generating structured function calls that are validated against JSON schemas before execution. The system supports native function-calling APIs from OpenAI, Anthropic, and other providers, with automatic routing to the most efficient provider based on function complexity and latency requirements. Function calls are type-checked and validated server-side before being passed to user code.

Follows complex, multi-part instructions with high fidelity by parsing instruction hierarchies and maintaining constraint satisfaction throughout generation. The model uses a constraint-aware decoding strategy that prevents violations of specified rules (e.g., 'respond in JSON only', 'use exactly 3 paragraphs', 'avoid mentioning X') by filtering the token probability distribution at each generation step to exclude tokens that would violate constraints.

Provides code completion and generation that understands the full context of a codebase by indexing function definitions, class hierarchies, and variable scopes. The model uses semantic search to retrieve relevant code snippets from the index and incorporates them into the context window, enabling completions that reference existing code patterns and maintain consistency with the codebase style and architecture.

Generates responses as a stream of tokens that can be consumed in real-time, with fine-grained control over token emission and the ability to stop generation early based on custom criteria. The streaming implementation uses a token queue that allows clients to inspect each token before it's sent, enabling use cases like token filtering, cost monitoring, and dynamic stopping based on semantic conditions (e.g., stop when a complete sentence is generated).

Learns from a small number of examples provided in the prompt (few-shot learning) by automatically selecting and ordering examples to maximize task performance. The model uses a learned ranking function to identify which examples are most relevant to the current task, and orders them to create an optimal learning trajectory where earlier examples establish patterns that later examples reinforce.

Generates content while enforcing safety policies and content guidelines through a multi-layer filtering system that operates at the prompt analysis, generation, and output stages. The model uses learned safety classifiers to identify potentially harmful requests, applies constraint-aware decoding to prevent unsafe content generation, and performs post-generation filtering to catch edge cases that bypass earlier layers.

Processes multiple requests in batches to optimize API costs and maximize throughput by grouping requests and processing them together. The batch system automatically schedules requests based on priority and deadline, packs them efficiently into API calls to minimize overhead, and applies cost-saving techniques like token deduplication and shared context caching across requests in the batch.