<p>OpenAI has released two new smaller models, GPT-5.4 mini and GPT-5.4 nano, aimed at high-volume and latency-sensitive workloads.</p>
<h2>What’s new</h2>
<p><strong>GPT-5.4 mini</strong> runs 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 supports tool use, function calling, multimodal inputs, and a 400k context window. Pricing is $0.75 per million input tokens and $4.50 per million output tokens.</p>
<p><strong>GPT-5.4 nano</strong> is OpenAI’s cheapest model to date at $0.20/$1.25 per million tokens. It outperforms GPT-5 mini at maximum reasoning effort and is suited for classification, data extraction, and simpler coding tasks. It’s API-only for now.</p>
<p>For context on the nano pricing: one developer ran the numbers and found you could describe every image in a 76,000-photo library for around $52 total - under 0.07 cents per photo. That’s cheaper than Google’s Gemini 3.1 Flash-Lite on input tokens.</p>
<h2>Where they’re available</h2>
<ul>
<li><strong>API</strong>: both models available now</li>
<li><strong>ChatGPT</strong>: mini available now</li>
<li><strong>Codex</strong>: mini available now, with a 0.33x premium request multiplier</li>
<li><strong>GitHub Copilot</strong>: mini rolling out now for Pro, Pro+, Business, and Enterprise users 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. Enterprise and Business admins need to enable it in Copilot settings first.</li>
</ul>
<h2>The subagent angle</h2>
<p>The timing isn’t accidental. 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 parallel focused tasks. Both small models are optimized for exactly this kind of role - fast, cheap, capable enough to handle a well-scoped subtask without needing the full model.</p>
<p>Anthropic and Google are pursuing similar strategies with Claude 4.5 Haiku and Gemini 3 Flash respectively, so this is less a unique architectural bet and more the industry converging on the same multi-model approach.</p>
<p>The <code>llm</code> CLI tool (v0.29) has already been updated to support both new models if you want to try them from the command line.</p>


Collections

OpenAI has released GPT-5.4 mini and GPT-5.4 nano, two smaller models targeting high-volume and latency-sensitive workloads. GPT-5.4 mini is over 2x faster than GPT-5 mini, supports tool use, function calling, multimodal inputs, and a 400k context window, priced at $0.75/$4.50 per million tokens. GPT-5.4 nano is OpenAI's cheapest model at $0.20/$1.25 per million tokens, outperforming GPT-5 mini at maximum reasoning effort and suited for classification, data extraction, and simple coding. Both are available via API now, with mini also in ChatGPT, Codex, and GitHub Copilot. The release aligns with OpenAI's new Codex subagent feature, where larger models handle planning while mini/nano instances run parallel focused tasks — a pattern also pursued by Anthropic and Google.

OpenAI releases GPT-5.4 mini and nano