OpenAI released two new models on March 17: GPT-5.4 mini and GPT-5.4 nano, both aimed at cost-sensitive deployments where running a full frontier model isn’t justified by the workload.
GPT-5.4 mini is the primary release. According to OpenAI’s internal evaluation, it came within 5% of GPT-5.4’s scores on SWE-Bench Pro and OS-World-Verified, two benchmarks focused on software engineering agents and operating system task completion. OpenAI also states the model completes some tasks more than twice as fast as GPT-5 mini. Neither benchmark result has been independently verified by Epoch AI or a third party at time of publication.
OpenAI states GPT-5.4 mini supports a 400,000-token context window with image input, per OpenAI’s announcement. Builder note: this figure and API pricing ($0.75 per million input tokens, $4.50 per million output tokens, per OpenAI’s announcement) appear in the source package but were not present in the verified page content retrieved for this cycle. Both figures carry qualified status and should be confirmed against the OpenAI official announcement before being cited as established facts in downstream content.
GPT-5.4 nano was released alongside mini as a more cost-optimized variant. Specific capability details for nano are not available in verified sources at time of publication.
The practitioner takeaway: if you’re running agentic workloads or long-document pipelines where full GPT-5.4 costs are prohibitive, GPT-5.4 mini warrants a direct evaluation. The 400,000-token context window, if confirmed, puts near-flagship context capacity into a materially cheaper wrapper. Verify pricing against OpenAI’s official announcement before building cost models.