arkennemasis Codex LLM (ChatGPT login)

ArkCodexLLM puts the GPT-5 family - text and vision - inside your ComfyUI graph, billed against your existing ChatGPT subscription instead of an API key. Same trick as the pack's Codex Image Gen: no key field, no in-UI OAuth, it reuses the login the Codex CLI wrote. It's the LLM node the Variation pipeline's "Step 1B" step expects, and it's a perfectly good general text/vision node for anything else in a workflow that wants a frontier model without paying per token.

The setup cost is one terminal command, once:

codex login

Requires a paid ChatGPT plan - the pack is developed against ChatGPT Plus at $20/month - and the CLI must be installed and signed in on the same machine and user account that runs ComfyUI. The node reads ~/.codex/auth.json (or $CODEX_HOME/a per-node codex_home), refreshes an expired token and saves it back atomically, and names the signed-in account in its account output.

The inputs that matter

• prompt - required; the user content.
• system_instructions - wire the pack's System Instructions node here, or type rules directly.
• image_1 through image_4 - optional vision inputs (can be batches). This is what makes it a VLM rather than a text box.
• model - the GPT-5 family, smartest first: gpt-5.6-sol, -terra, -luna, then 5.5, 5.1, 5. The tooltip is honest: availability depends on your plan; if one is refused, try the next. model_override takes a raw model id for anything newer than the dropdown.
• reasoning_effort - default through max. The tooltip's guidance is the good stuff: ultra is the Codex CLI's own workhorse and worth it for story planning, where one call decides the quality of everything downstream. default omits the field entirely.
• json_only - demand bare JSON and strip ``` fences. Turn it on whenever a downstream node parses the text; the pack's Recipe Compile step is exactly this shape, and an LLM answering with a code fence is how a pipeline breaks at 2am.
• batch_total + batch_size - split a long structured answer across several calls. A 16-scene plan is ~12,000 output tokens in one response, which is slow, goes silent long enough for the connection to cut, and can hit the output cap and come back truncated. Set batch_total with batch_size (default 5) and earlier items are passed forward so the story still joins up. This is the single most useful setting on the node for real work.
• speed - fast sends service_tier=priority (the CLI's "1.5x speed, increased usage"). Measured on the author's account: first token 0.75s vs 1.00s. It spends plan quota faster, so leave it standard for long batch runs.

Outputs: text (STRING) and account. Note GPT-5 models output text only - image generation is the Codex Image Gen node's job.

Installing and gotchas

Part of comfyui-arkennemasis:

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis
pip install replicate httpx

restart, under arkennemasis/LLM. Like all the Codex nodes, it needs httpx to load and a live CLI login to run.

Where people get burned: long single-call JSON plans (use batching - it's there for exactly the truncation you'll hit), and forgetting that one at a time run_mode serialises every arkennemasis API node in the graph. If your image gens are queueing behind a slow planning call, that's the lock doing its job; give the planner its own graph or accept the serialisation.