arkennemasis Replicate LLM (OpenAI GPT-5)

An LLM node in a ComfyUI graph is a tool, not a component of the model - it writes your structured prompts, drafts story plans, captions images, and feeds text to the rest of the workflow. This node is the arkennemasis pack's hosted option: an OpenAI GPT-5 family model served on Replicate, with text and vision (up to four image inputs). It's pay-as-you-go per call, so the honest framing is exactly the one the KB gives for this whole category: reach for it when you want genuine large-chat quality and you don't need local or uncensored - anything the API filters, this node can't help with.

What you get

Two required inputs: model and prompt. The model dropdown is the full GPT-5 family - openai/gpt-5, -mini, -nano, -pro, -structured, plus 5.1, 5.2, 5.4 and the 5.6-luna/terra/sol variants. The single output is text (STRING).

The optional inputs carry the interesting parts:

• system_prompt - type it here, or wire in the pack's System Instructions node.
• image_1 through image_4 - vision inputs, each accepting a batch. This is what makes the node a VLM: wire a reference image in and the model reads it.
• reasoning_effort (default/minimal/low/medium/high) and verbosity (default/low/medium/high) - both left on default are not sent, so the model's own defaults apply. That sentinel pattern runs through the pack.
• max_completion_tokens - 0 = model default.
• api_token - Replicate key; blank falls back to REPLICATE_API_TOKEN env var or a .env file (pack folder or ComfyUI root).
• timeout_seconds - 0 = wait indefinitely.
• force_rerun - re-call the API even if inputs are unchanged.
• run_mode - one at a time (default) serialises every arkennemasis API node in the graph via an asyncio lock; all at once runs them concurrently up to max_concurrent (default 2, 0 = uncapped). The default exists because ComfyUI runs async nodes concurrently, and Replicate drops to a burst of 1 under $5 credit - parallel calls then reliably 429.

Behind the scenes it polls the prediction off the UI thread, so ComfyUI stays responsive and Cancel works, and it retries the failure modes that deserve retrying - 429s with backoff, 5xx, dropped connections - while never retrying a moderation refusal or an auth failure.

The workflow shape

The pack's canonical flow: System Instructions → this node (system_prompt) with your text and any reference images → its text output feeds the Image Gen node's prompt. It's also the natural story-agent node for the video pipeline - the Story Brief node's output wires straight into prompt.

Install

ReplicateOpenAILLM is one of the 61 nodes in the comfyui-arkennemasis pack, in arkennemasis/LLM:

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis
pip install -r comfyui-arkennemasis/requirements.txt   # then restart ComfyUI

Then get a token at https://replicate.com/account/api-tokens and either paste it into the node, set REPLICATE_API_TOKEN, or add it to a .env in ComfyUI's root (kept outside the repo so re-cloning never touches it). Or use ComfyUI Manager → Install via Git URL with the repo URL.

The one thing to remember: every call bills your Replicate account. If you already pay for ChatGPT, the pack's Codex LLM path is the same idea without per-call billing - worth checking before you spend.