OpenAI on March 17, 2026, released GPT-5.4 mini and GPT-5.4 nano, two smaller models in its GPT-5.4 family designed for speed, lower cost and high-volume automated workloads. The company says the mini model runs about twice as fast as its predecessor, GPT-5 mini, while the nano model is the lightest and cheapest option in the lineup.
Both models are available immediately through the OpenAI API, in the Codex coding tool and, for the mini model, as a fallback option within ChatGPT when usage limits on larger models are reached.
Built for coding and subagents
OpenAI positions the models for tasks where many calls happen in quick succession: coding assistance, computer-use agents that operate software on a user’s behalf, and so-called subagent workloads, in which a larger model delegates narrower jobs to smaller, cheaper models running in parallel. According to OpenAI’s launch materials as reported by The New Stack and The Decoder, GPT-5.4 mini scores close to the full GPT-5.4 model on several coding and agentic benchmarks at a fraction of the latency.
The models are multimodal, accepting images as well as text, and support the same tool-calling and structured-output features as the larger GPT-5.4 model, which lets developers swap them into existing pipelines without rewriting integrations.
Faster, but not cheaper than the last generation
One detail that drew attention in developer circles is pricing. The Decoder reported that GPT-5.4 mini and nano cost up to four times more per token than the GPT-5 mini and nano models they succeed, reflecting the step up in capability. OpenAI’s argument is that higher per-token prices are offset by fewer retries and shorter outputs, but the change complicates the usual expectation that each generation of small models gets cheaper.
For developers, the practical question is whether the speed and accuracy gains justify the higher bill for a given workload, particularly for simple classification or extraction tasks where the older models performed adequately.
Context
The release continues a pattern in which OpenAI pairs each flagship model with smaller siblings for cost-sensitive use. GPT-5.4 itself arrived earlier in March 2026 with an emphasis on long-running agent tasks and computer use. Competitors offer similar tiers: Anthropic’s Haiku line, Google’s Gemini Flash and Flash-Lite models, and a growing set of open-weight models from Chinese labs that undercut all of them on price.
Early reactions from developers, as collected in trade coverage, were generally positive on speed and coding quality, with the pricing shift the main point of criticism. Independent benchmark results beyond OpenAI’s own figures were still emerging at the time of the launch.
