GPT-5.4 Nano

Some LLM jobs are boring on purpose: "is this caption about a landscape?", "extract the three words that describe the style", "clean up this prompt's commas." For that kind of high-volume, low-stakes text work, you do not want the big model - you want the smallest thing that reliably answers. GPT-5.4 Nano is exactly that: the cheapest, quickest tier of the 5.4 family, with the full control surface intact.

How it works

A textInference node that calls GPT-5.4 Nano through the Runware cloud API. The required messages input is a RUNWARE_MESSAGES socket, fed by the Runware Messages builder (Runware/Params) - same pattern as every other text node in the pack.

The controls are identical to the full 5.4 node, which is the convenient part:

• settings.systemPrompt - system instruction; on Nano this is where you keep the task tight ("answer only yes or no").
• settings.temperature (default 1, 0–2) and settings.topP (default 0.95) - you'll probably run Nano at a low temperature for consistent answers.
• settings.thinkingLevel - (default)/low/medium/high/xhigh. Nano rewards low; it's not built for deep reasoning, so don't pay for thinking it can't really do.
• settings.maxTokens (default 4096, up to 128k) - response cap.
• images (IMAGE) - yes, even Nano is multimodal; handy for quick "does this image show X?" checks.
• toolChoice / toolChoice.type / toolChoice.name and advanced_json for tools, plus seed, numberResults, includeUsage.

Output is a text STRING.

Install

Ships in ComfyUI-Runware:

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

Restart (or ComfyUI Manager → search "Runware"). The model runs in the cloud, so there's nothing to download. API key: Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Gotchas

Nano is the value pick but it's also the one people misjudge. It will confidently produce mediocre results on anything that needs real reasoning - use it for classification and extraction, not for "improve my prompt into a masterpiece." Per-token cost is the lowest of the family, but it's still per-token; the title bar shows each run's cost. And as with every text node here, no chat memory - stateless per execution, multi-turn is your job to wire up.