OpenAI launches GPT-5.4 mini and nano for faster coding and subagent workloads

OpenAI expanded its small-model lineup on March 17, 2026 with GPT-5.4 mini and GPT-5.4 nano. The company describes them as its most capable small models yet, built to carry many of the strengths of GPT-5.4 into faster and more efficient deployments. The product framing is straightforward: for high-volume workloads, the most useful model is often not the largest one, but the one that can respond quickly, use tools reliably, and stay competent on professional tasks.

The flagship of the announcement is GPT-5.4 mini. OpenAI says it runs more than 2x faster than GPT-5 mini while improving performance across coding, reasoning, multimodal understanding, and tool use. The benchmark table is strong enough to explain the launch on its own. On SWE-Bench Pro (Public), GPT-5.4 mini scored 54.4% versus 45.7% for GPT-5 mini. On Terminal-Bench 2.0 it reached 60.0% versus 38.2%. On Toolathlon it posted 42.9% versus 26.9%, and on OSWorld-Verified it came in at 72.1% versus 42.0%. OpenAI also says mini approaches the larger GPT-5.4 model on several evaluations, including SWE-Bench Pro and OSWorld-Verified.

Small models as operating units for agent systems

What matters here is not just the raw improvement over GPT-5 mini, but how OpenAI positions the model inside larger systems. The post explicitly describes Codex workflows where a larger model handles planning, coordination, and final judgment, while GPT-5.4 mini subagents take narrower tasks in parallel, such as searching a codebase, reviewing a large file, or processing supporting documents. That language makes mini look less like a cheap fallback and more like an operational building block for agentic software workflows.

GPT-5.4 nano pushes the same strategy further toward price and latency. OpenAI says it is the smallest and cheapest GPT-5.4 model and recommends it for classification, data extraction, ranking, and simpler coding subagents. Even so, it is not framed as a minimal utility model. The company reports 52.4% on SWE-Bench Pro (Public), 35.5% on Toolathlon, and 82.8% on GPQA Diamond. That gives developers a lower-cost option for supportive tasks that still require real coding, tool use, and reasoning competence.

Availability and pricing

Availability is broad for mini and selective for nano. GPT-5.4 mini is available immediately in the API, Codex, and ChatGPT. In the API it supports text and image inputs, tool use, function calling, web search, file search, computer use, and skills, with a 400k context window. OpenAI prices it at $0.75 per 1M input tokens and $4.50 per 1M output tokens. In Codex, the model uses only 30% of the GPT-5.4 quota. GPT-5.4 nano is API-only and costs $0.20 per 1M input tokens and $1.25 per 1M output tokens.

The broader signal is that OpenAI is segmenting its model stack around real product roles. Larger models still anchor planning and difficult reasoning, but smaller models are being optimized to keep coding assistants responsive, to interpret screenshots for computer use, and to execute subagent work at scale. This release is a concrete step toward multi-model AI systems where speed, cost, and tool competence matter as much as absolute frontier capability.