OpenAI: GPT-5 Mini

GPT-5 Mini is a text model from OpenAI, released on 7 August 2025. It is a compact version of GPT-5, built to handle lighter-weight reasoning tasks. According to OpenAI, it keeps the same instruction-following and safety-tuning benefits as the full GPT-5 model, but with reduced latency and cost. It sits in the text category and is aimed at developers who want GPT-5 style behaviour without the weight of the larger model.

The model accepts text, images and files as input, and returns text only. It has a context window of 400,000 tokens and can produce up to 128,000 output tokens, which is a large budget for long generations such as multi-page drafts or extended structured output. Its knowledge cutoff is 31 May 2024, so anything after that date has to come from the prompt, from tools, or from retrieved context rather than from the model's own training. Reasoning is a first-class part of the interface: the supported parameters include reasoning, reasoning_effort, include_reasoning and verbosity, so you can tune how much internal reasoning the model performs and how much of it is surfaced.

Those capabilities point to a fairly wide set of practical uses. Because it handles images and files alongside text, it fits document question answering, extraction from PDFs or screenshots, and summarisation of mixed content. The large context window suits long conversations, codebase review, and analysis of lengthy reports where you want the whole document in one request. Tool calling and structured outputs make it a reasonable choice for agents, function-calling workflows, and pipelines that need JSON or schema-conformant responses. The reasoning controls let you dial effort up for harder problems and down for simple classification or routing, which is useful when you want predictable behaviour across a mixed workload.

You call it through api14's OpenAI-compatible API using the model id "openai/gpt-5-mini". The supported request parameters are include_reasoning, max_completion_tokens, max_tokens, reasoning, reasoning_effort, response_format, seed, structured_outputs, tool_choice, tools and verbosity. In practice that means you can set a token cap, request a specific response format or schema, control tool use, fix a seed for reproducibility, and adjust reasoning effort and verbosity in the same request. Since the surface is OpenAI-compatible, existing OpenAI SDK clients and code generally work with only the base URL and model id changed.

One thing to keep in mind is the knowledge cutoff of 31 May 2024, so tasks that depend on recent events need retrieval or tools rather than the model's memory. It is also described as a compact model for lighter-weight reasoning, so if your workload is dominated by the heaviest reasoning problems, a larger model in the same family may be the better fit. For everything else, the combination of a 400,000 token context window, image and file input, tool calling and structured outputs covers a lot of ground.