OpenAI: GPT-4 Turbo (older v1106)

Processes both text and image inputs simultaneously within a single inference pass, using a unified transformer architecture that encodes visual tokens alongside text embeddings. The model applies attention mechanisms across both modalities, enabling it to reason about image content, answer questions about visual elements, and generate text responses grounded in visual context. Vision inputs are converted to image tokens through a learned visual encoder before being fed into the main language model backbone.

Constrains model output to valid JSON matching a developer-provided schema, using a decoding-time constraint mechanism that prevents invalid JSON generation at the token level. The model's output is validated against the schema before being returned, ensuring type correctness, required field presence, and enum constraints. This works by modifying the sampling distribution at each token position to only allow tokens that keep the output valid JSON.

Accepts a list of tool/function definitions with parameters, and the model learns to emit structured function calls in response to user queries. The model outputs function names and arguments as JSON, which the developer's application then executes and feeds back to the model for continued reasoning. This enables agentic workflows where the model decides which tools to invoke, in what order, and how to interpret results. The model is trained to understand function signatures, parameter types, and return values.

Processes input sequences up to 128,000 tokens (approximately 96,000 words or 400+ pages of text) in a single request, enabling the model to maintain coherent reasoning across very long documents, codebases, or conversation histories. The model uses a modified attention mechanism (likely sparse or hierarchical attention) to handle the extended context efficiently without quadratic memory scaling. This allows developers to pass entire books, code repositories, or long conversation threads without truncation.

Interprets natural language instructions and system prompts to adapt behavior without fine-tuning, using in-context learning to understand task specifications from examples (few-shot) or descriptions (zero-shot). The model's training includes extensive instruction-following data, enabling it to understand complex, multi-step tasks described in plain English and execute them consistently. This works through the model's learned ability to parse instructions, extract intent, and apply that intent to new inputs.

Generates syntactically correct code across 40+ programming languages (Python, JavaScript, Java, C++, Go, Rust, etc.) based on natural language descriptions, comments, or partial code. The model understands language-specific idioms, standard libraries, and best practices for each language. Code generation works through transformer-based sequence-to-sequence prediction, where the model learns patterns from billions of tokens of code in its training data and predicts the most likely next tokens that form valid code.

Explicitly acknowledges its training data cutoff (April 2023) and can reason about what information it may not have access to, enabling developers to build systems that know when to query external data sources. The model understands temporal references in queries and can indicate uncertainty about recent events or developments. This is implemented through training data that includes explicit temporal markers and examples of the model declining to answer about post-cutoff events.

Solves mathematical problems including algebra, calculus, geometry, and logic through step-by-step reasoning, using chain-of-thought patterns learned during training. The model can work through multi-step problems, show intermediate steps, and explain reasoning. This works by training the model on mathematical problem-solving datasets and using reinforcement learning to reward correct final answers and clear reasoning paths. The model learns to recognize mathematical patterns and apply appropriate solution strategies.