<p>OpenAI has released two new smaller models: GPT-5.4 mini and GPT-5.4 nano. Both are designed for high-volume, latency-sensitive work, particularly as subagents inside larger agentic pipelines.</p>
<h2>What they can do</h2>
<p>GPT-5.4 mini is more than 2x faster than GPT-5 mini and scores within 3 percentage points of the full GPT-5.4 on SWE-Bench Pro and OSWorld-Verified. It’s optimized for coding, computer use, multimodal understanding, and running parallel subtasks. GPT-5.4 nano is the cheaper option, better suited for classification, data extraction, and simpler coding subagents. It outperforms the previous GPT-5 mini at maximum reasoning effort despite being the smallest model in the lineup.</p>
<p>Both support tool use, function calling, multimodal inputs, and a 400k context window.</p>
<h2>Pricing</h2>
<ul>
<li><strong>GPT-5.4 mini:</strong> $0.75/1M input tokens, $4.50/1M output tokens</li>
<li><strong>GPT-5.4 nano:</strong> $0.20/1M input tokens, $1.25/1M output tokens</li>
</ul>
<p>The nano pricing undercuts Google’s Gemini 3.1 Flash-Lite. To put it concretely: describing a single museum photo costs under 0.07 cents, which means processing 76,000 photos would run about $52 total.</p>
<h2>Where they’re available</h2>
<p>Both models are available via the API. GPT-5.4 mini is also accessible in ChatGPT and Codex. It’s rolling out in GitHub Copilot as well, with the fastest time to first token among mini models and improved codebase exploration. Copilot Pro, Pro+, Business, and Enterprise users can access it across VS Code, Visual Studio, JetBrains, Xcode, Eclipse, <a href="http://GitHub.com" target="_blank" rel="noopener nofollow">GitHub.com</a>, GitHub Mobile, and GitHub CLI, though a 0.33x premium request multiplier applies. Enterprise and Business admins need to enable the model in Copilot settings first.</p>
<h2>The subagent angle</h2>
<p>The timing is deliberate. OpenAI launched subagents in Codex the day before releasing these models. The intended pattern is straightforward: a larger model handles planning and review while mini or nano instances run focused tasks in parallel. Anthropic and Google are pursuing similar strategies with Claude 4.5 Haiku and Gemini 3 Flash.</p>
<p>The <code>llm</code> CLI tool (v0.29) has been updated to support both new models.</p>


Collections

OpenAI has released GPT-5.4 mini and GPT-5.4 nano, two small models optimized for high-volume, latency-sensitive agentic workloads. GPT-5.4 mini is over 2x faster than GPT-5 mini and scores close to the full GPT-5.4 on coding benchmarks, while nano is the cheapest option suited for classification and simple subagent tasks. Both support tool use, function calling, multimodal inputs, and a 400k context window. Pricing is $0.75/$4.50 per million tokens for mini and $0.20/$1.25 for nano, undercutting Google's Gemini 3.1 Flash-Lite. Both are available via API; mini is also in ChatGPT, Codex, and rolling out in GitHub Copilot across major IDEs. The release is timed with OpenAI's Codex subagent launch, positioning these models as worker nodes in multi-agent pipelines alongside competitors like Claude 4.5 Haiku and Gemini 3 Flash.

OpenAI releases GPT-5.4 mini and nano