Anthropic Claude

Anthropic's Claude is the model people reach for when they want careful long-form reasoning and strong instruction-following, and this node drops it straight into your node graph. Prompt in, STRING out, with up to 20 images as vision context. It's the partner/text node with the most grown-up knobs: per-model reasoning budgets, token caps, and refusal handling that tells you when the model declined instead of pretending.

The mechanism

The node talks to Claude's Messages API through Comfy's proxy and bills through your Comfy account in credits - no Anthropic key of your own, price badge on the node, one of the family's predictable catches. Images you attach get uploaded to Comfy storage and passed as URLs, downscaled to a fixed pixel budget so a 4K render doesn't blow up the request. It landed in ComfyUI core in May 2026.

The inputs that matter

• prompt - the text, multiline.
• model - a dynamic combo over the current Claude lineup (Opus 5, Opus 4.8, Fable 5, Sonnet 5, the 4.6/4.7 generation, Haiku 4.5). Pick one and the node reveals that model's controls: max_tokens (default 32768, up to 64000), temperature, and reasoning_effort for extended thinking. The dropdown updates as Anthropic ships new models - the newest ones drop the temperature knob entirely because they don't accept it.
• seed - same caveat as the rest of the family: it gates re-runs, not determinism.
• images - the autogrowing image list, capped at 20. That's a hard limit; go over and the node errors out.
• system_prompt - optional foundational instructions.

Output is a STRING. Wire it anywhere text goes.

Getting it and what to watch

Ships with ComfyUI core, nothing to install. The interesting failure modes are Claude-specific. If the model refuses for safety reasons, the node surfaces that explicitly - "Claude declined to answer this request" - rather than returning silent garbage, so rephrase or switch models. Extended thinking is the big lever: reasoning_effort burns thinking tokens inside max_tokens, so if you crank reasoning and responses come back empty or truncated, raise max_tokens (the budget logic reserves room for the actual answer, but only up to a point). And it's cloud-only with Comfy credits - same storefront model as every node in this family. Good node, opinionated pricing, excellent if you need Claude's reasoning inside a pipeline that's already on Comfy's cloud.