GPT-5.4 Pro

Some problems in a workflow deserve a model that will sit there and think for a while: "design a regional prompt setup that fixes the hands in this generation chain" or "explain why this lighting pipeline keeps banding." That's what GPT-5.4 Pro is for. It's the biggest, most capable tier of the 5.4 family on the Runware pack, and you'll know it when you see the bill - which is exactly why it's a specialist, not a default.

How it works

A textInference node calling GPT-5.4 Pro through the Runware cloud. The required messages input is a RUNWARE_MESSAGES socket from the Runware Messages builder (Runware/Params), same as the other text nodes.

Here's the interesting part about the Pro's control surface: there's no temperature or topP on this node. OpenAI locks those on the big reasoning models - you don't get to dial creativity, the model's internal sampling is what it is. What you do control:

• settings.thinkingLevel - (default)/low/medium/high/xhigh. This is the main lever, and on Pro it's the whole game: crank it when the problem is genuinely hard. It's also where your budget goes.
• settings.systemPrompt - the system instruction.
• settings.maxTokens (default 4096, up to 128k) - response cap; long reasoning chains eat into this.
• images (IMAGE) - Pro is multimodal, so it can stare at your actual images before advising on them. This is the killer feature for "look at my render and tell me what's wrong."
• toolChoice / toolChoice.type (auto/any/tool/none) and toolChoice.name, with advanced_json for tool definitions.
• numberResults (up to 4) and includeUsage.

Output is a single text STRING.

Install

Part of 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"). Cloud model, no downloads. API key: Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Gotchas

Cost, cost, cost. Pro with xhigh thinking on a long prompt is the most expensive single node run in the text category - the title bar shows the damage, but glance before you queue a batch. And don't be surprised by the missing temperature/topP; that's OpenAI's policy on reasoning models, not a pack bug. If you find yourself reaching for it on trivial jobs, you're paying for a specialist to do grunt work - the Mini or Nano handles those.