Claude Opus 4.5

Here's the thing that confuses people about this pack: there's a separate Opus 4.5 node and a 4.6, 4.7, 4.8 and 5. That's not clutter, it's the design. KIE publishes a documentation page per model version, and this pack builds one ComfyUI node per published page. Retired-ish versions stay callable, which is exactly what you want when the workflow you shipped six months ago is pinned to the version you validated against.

So: open this node when you're maintaining an old graph, reproducing an old result, or deliberately freezing a prompt pipeline so a model refresh doesn't quietly change your output. For new work, start at the newest Opus and walk backwards only if something breaks.

What it's for

Same job as every LLM node here - text in, text out, upstream of the sampler. Prompt rewriting, structured scene description, turning rough notes into a format your video model likes. It is not a text encoder; if you're hazy on that, read llm-in-comfyui.md first, because conflating the two is the most common way people get confused about what an API LLM is even doing in their graph.

The honest pitch for this specific node: version pinning. A prompt enhancer is a dependency. When a provider refreshes a model, the same prompt can start returning a different dialect - more flowery, more refusal-prone, longer. If your workflow's output depends on the LLM's phrasing, having 4.5 still sitting in the menu is a genuine feature and one of the more thoughtful things about this pack.

How it works

The node class is generated at startup from KIE's published Claude docs, and it builds an Anthropic-shaped request - model, messages, stream, max_tokens, optional system, tools and thinkingFlag - posted to /claude/v1/messages on KIE's API domain. KIE is a reseller: you're paying KIE's marketplace rate to reach Anthropic's model, and the practical reason people do that is price. One community pack author measured KIE's Nano Banana Pro at roughly $0.09–0.12 per image versus $0.15–0.30 via Comfy's native API node, and reported throwing $10 at KIE and generating a lot of content with it.

history_json is passed through as the messages array with your prompt appended last. thinking maps to KIE's thinkingFlag, not Anthropic's budgeted thinking object.

Inputs and outputs

Required: prompt. Optional: system_prompt (put your output contract here - "return one paragraph, no headings"), stream, tools_json, history_json, max_tokens (default 4096, up to 131072), thinking, expert_override_json.

Outputs: text is what you wire downstream into a CLIPTextEncode or another model's prompt input. raw_json is the full response - the thing you go read when the output looks mangled. usage_json, credits_consumed and credits_left are numbers/JSON you'll mostly use for accounting; the node also shows a spent/left receipt on its face after each run.

No IMAGE input on Claude nodes in this pack, by design.

Install

Manager, search ComfyUI-KIE-Nodes-Next. Registry publishing is prepared but not shipped as far as the README goes, so the manual path is:

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
cd ComfyUI-KIE-Nodes-Next
pip install -r requirements.txt

Restart, then Settings → KIE.ai Nodes Next → Connection and paste your key (validated before saving, shown only as a masked fingerprint). KIE_API_KEY takes priority for headless installs. Python 3.10+, no GPU, no weights.

Common issues

Not in the node list yet. Startup builds the library from KIE's live docs and reloads the UI once. Restart and be patient before you file anything.

The workflow you built in the web UI doesn't match. Old node classes are stable by design; if a version was dropped from KIE's catalog, the endpoint call can fail with a provider-side error even though the node loads fine. The pack's refresh-and-retry behaviour is about the catalog, not about a model that no longer exists upstream.

Empty output, valid API call. Usually max_tokens too low for the JSON or long prompt you asked for, or a system_prompt that never told it to skip the preamble.

This pack has almost no community footprint. Searching the corpus finds zero threads naming it - the Kie.ai node conversation out there is about a different pack. That doesn't make it bad code (the key handling is unusually careful for this category), but it does mean nobody has stress-tested it for you, and it holds a credential by design. Read before you install, and don't reuse a key you care about elsewhere.