OpenAI: GPT-5 Nano

GPT-5 Nano is the smallest and fastest model in the GPT-5 family, optimised for ultra-low-latency environments — developer tooling, rapid interactions, autocomplete-style features and high-throughput pipelines. It is the successor to GPT-4.1-nano and, like the rest of the family, retains GPT-5-generation instruction-following and safe-completions behaviour rather than older response styles.

Its defining characteristic is price-performance at scale. Nano costs roughly $0.05 per million input tokens and $0.40 per million output tokens — the cheapest tier of the GPT-5 system — while still offering the 400,000-token context window. That combination makes it viable to run GPT-5-generation behaviour across very large volumes of requests that would be uneconomical on the larger tiers.

Nano deliberately trades reasoning depth for speed and cost. It will not match full GPT-5 or even Mini on hard multi-step problems, and it benefits more than the larger tiers from clear prompting, good retrieval and well-structured tasks. The right mental model is a fast, cheap component you compose into a larger system — not a standalone solver for your most demanding work.

In a routed architecture, Nano is the floor: the tier you reach for when latency and unit economics matter most and the task is narrow and well-defined. Pair it with an escalation path so that requests it cannot handle confidently are routed up to Mini or full GPT-5, ideally driven by your own quality checks rather than fixed rules.

Nano fits real-time and high-volume scenarios: inline suggestions and autocomplete, lightweight classification and tagging, fast intent detection, simple transformations, and the cheapest steps inside an agent pipeline. Its latency makes it suitable for features embedded directly in an editing or typing experience, where a slower model would feel sluggish.

Avoid Nano for tasks that require sustained reasoning, nuanced judgement or broad world knowledge under ambiguity — those belong to Mini or full GPT-5. Used well, Nano is the model that lets you apply GPT-5-generation behaviour to parts of your product that would otherwise be too high-volume or latency-sensitive to justify a larger model.