OpenAI Launches GPT-5.4 Mini and Nano for Speed-First Workloads

TL;DR:

• OpenAI has released GPT-5.4 mini and nano, smaller models designed for coding, subagent orchestration, and computer use tasks where speed matters more than raw capability
• GPT-5.4 mini runs over twice as fast as its predecessor while scoring 54.4% on SWE-Bench Pro, approaching the full GPT-5.4’s 57.7%
• The release signals a broader industry shift towards multi-model architectures, where smaller specialised models handle routine tasks under the direction of larger ones

OpenAI has introduced two new models to its GPT-5.4 family: mini and nano. Rather than pushing for maximum capability, these models target the growing demand for fast, cost-efficient AI that can handle high-volume workloads without the overhead of full-scale reasoning models.

Built for Developer Workflows

GPT-5.4 mini represents a substantial step up from GPT-5 mini across coding, reasoning, and tool use benchmarks. On SWE-Bench Pro, it reaches 54.4% compared to GPT-5 mini’s 45.7% — a nearly 20% relative improvement. The model is priced at $0.75 per million input tokens and $4.50 per million output tokens, with a 400k context window.

The nano variant sits even lower on the cost curve at $0.20 per million input tokens, targeting classification, data extraction, and simple coding subtasks. It remains API-only for now.

The Multi-Model Future

Perhaps the most telling aspect of this release is OpenAI’s emphasis on subagent architectures. The company explicitly recommends using GPT-5.4 mini as a worker model directed by the larger GPT-5.4, with the bigger model handling planning and coordination while mini executes narrower tasks in parallel. This pattern — already visible in OpenAI’s Codex product — reflects a growing consensus across the AI industry that production systems will increasingly combine models of different sizes rather than relying on a single large model.

Looking Forward

For UK businesses building AI-powered products, the practical takeaway is clear: the cost of deploying capable AI continues to fall. GPT-5.4 mini in Codex uses just 30% of the standard quota, making sophisticated coding assistance more accessible for smaller development teams. As the gap between “big” and “small” models narrows on practical benchmarks, the economics increasingly favour multi-model approaches — a trend that could reshape how organisations budget for and architect their AI systems.