OpenAI Releases GPT‑5 with Reasoning Router and 400 k‑Token Context

OpenAI has unveiled GPT-5, described as its smartest, fastest, and most useful model so far. In ChatGPT, GPT-5 is now the default model, replacing GPT-4o, OpenAI o3, o4-mini, GPT-4.1, and 4.5. For tougher work, GPT-5 Pro (“GPT-5 thinking pro”) offers extended test-time reasoning for the most demanding tasks.

One unified system with a real-time router

Rather than a single monolith, GPT-5 is a unified system combining two “brains”: a fast model for everyday queries and a deep reasoning model for harder problems. A real-time router decides which to use based on task complexity, tool requirements, and the user’s intent (e.g., “think deeply about this”). The goal is to deliver expert-level answers when needed while keeping day-to-day interactions snappy. One might joke that GPT-5 has both a sprinter brain and a marathon brain working in tandem behind the scenes.

Long context and developer controls

On the API side, GPT-5 ships in three sizes — gpt-5, gpt-5-mini, and gpt-5-nano. It supports a total 400,000-token context (about 272k input and 128k for reasoning/output). Two new controls help tailor behavior and cost: reasoning_effort (to minimize or increase the model’s reasoning effort) and verbosity (short, medium, or long responses). OpenAI also added custom tools (free-form tool calls constrained by a grammar) to make the agent more reliable during long, multi-step action sequences.

Performance highlights

OpenAI reports strong gains in coding, math, writing, and health. For software engineering, GPT-5 reaches 74.9% on SWE-bench Verified and 88% on Aider Polyglot, with improved efficiency (fewer tokens and tool calls) compared to prior models. In agentic tool use, it hits around 97% on τ²-bench telecom. For long-context tasks, internal tests show high accuracy up to 128–256k tokens, and the model remains effective at retrieving information even from very large inputs.

In writing and creativity, GPT-5 handles structural ambiguity better (e.g., free verse poetry or strict stylistic constraints), yielding more nuanced, rhythmic prose. In health, it scores higher on realistic evaluations (while not replacing clinicians), acting more like a thought partner that proactively asks clarifying questions and flags uncertainties.

Reliability and safety

OpenAI highlights a meaningful reduction in hallucinations versus GPT-4o and earlier reasoning models, especially when web browsing is available. GPT-5 has been trained to admit its limits (better at abstaining or acknowledging impossible requests) and to reduce sycophancy toward user opinions. On safety, GPT-5 Thinking is treated as a High capability system for bio/chem domains, with layered safeguards: safe-completions, dedicated monitors, external red teaming, and a strengthened preparedness framework.

Availability

In ChatGPT, GPT-5 is live for all signed-in users (with higher usage limits on Plus/Pro; Pro subscribers also unlock GPT-5 Pro). On the API, developers can pick gpt-5 / mini / nano to balance cost, latency, and quality as needed.

Bottom line: GPT-5 marries a reasoning router, a massive 400k-token context, and tighter API controls to deliver practical gains in coding, agentic workflows, writing quality, and reliability.