OpenAI released GPT-5.4 Mini and GPT-5.4 Nano on March 17, 2026, two weeks after launching the flagship GPT-5.4 model on March 5.
The releases extend the GPT-5.4 family across a range of performance and cost profiles, enabling deployment strategies where different model sizes handle differentiated workloads within applications. GPT-5.4 Mini is priced at $0.75 per million input tokens and $4.50 per million output tokens, while the more compact Nano costs $0.20 per million input tokens and $1.25 per million output tokens. Mini runs at twice the speed of the previous GPT-5 Mini model.
Performance benchmarks demonstrate that the smaller models retain significant capability. GPT-5.4 Mini scored 54.4 percent on SWE-Bench Pro, a benchmark measuring software engineering tasks, and 72.1 percent on OSWorld-Verified, which tests instruction-following on real operating systems.
These results trail the flagship model’s 75.0 percent on OSWorld but represent substantial capability for cost-optimized inference. GPT-5.4 Nano, despite its smaller size, outperforms the previous GPT-5 Mini model even at maximum reasoning effort, suggesting that the improvements from the GPT-5.4 generation apply across the family.
Use case targeting indicates where OpenAI expects the different models to serve. Mini is recommended for applications requiring coding, reasoning, multimodal understanding, and tool use. Nano is optimized for classification, data extraction, ranking, and agent subcomponents. Mini is available through both the API and Codex platforms; Nano is available through API only. OpenAI integrated GPT-5.4 Mini as the default model powering ChatGPT’s free tier, making frontier-level coding and reasoning accessible to users without subscription.
The flagship GPT-5.4 model brings native computer-use capabilities to a mainline model for the first time, along with 1 million token context windows and the 83 percent GDPVal score indicating performance on economically valuable tasks. The multi-model release strategy allows developers to architect systems where smaller models handle routine tasks and larger models address higher-complexity problems, a pattern OpenAI frames as enabling the “subagent era” of AI deployment.
