Gemini 2.5 Pro (openai)

Pro-tier models inside a graph are for the jobs where Flash isn't enough: planning a multi-shot sequence from a one-line idea, reformatting a messy transcript into structured JSON, or holding a long brief in context while you interrogate it. 2.5 Pro is Google's older reasoning tier, still on KIE's catalog, and still a sensible pick when it's the cheapest Pro-tier option available or when you have a workflow pinned to an API version that already behaves the way you expect.

If you're choosing fresh, the 3.x Pro node in the same folder is the newer tier. If you're here because the number looked familiar, that's fine too - the node shape is identical, so nothing about your graph changes when you swap.

How it works

One chat call to KIE's OpenAI-compatible Gemini route, with your prompt turned into a messages array and the response parsed down to a single string. (openai) in the name means OpenAI-schema request, not Gemini-native - which matters if you're pasting tool definitions or a payload example you got from an OpenAI doc, because that's the schema this endpoint speaks.

No image socket, no samplers, no VRAM. The whole node is prompt in, text out, plus token and credit accounting.

Inputs

prompt - required. The user message. For planning work, put the whole brief here and ask for the output format you want explicitly.

system_prompt - the job contract. An LLM won't reliably deviate from its chatty habits without one. If you want raw JSON back and nothing else, that instruction lives here.

history_json - a prior-messages array passed through untouched. This is how you get a conversation out of a node that has no chat UI: keep the array in your own flow, feed it back each run. It's also the only route to anything multimodal on this node, since the pack doesn't declare an image input for the Gemini family - the array goes to the endpoint as-is, and whether it accepts a hand-written image content part is between you and the API.

tools_json - OpenAI-style function definitions, [] if unused. The docs page's examples include a plain googleSearch tool entry.

stream - off by default. The pack collects a streamed response back into one string, so this is a transport detail rather than live typing in the UI.

expert_override_json - the raw request override for fields KIE's docs didn't publish. Normal use doesn't need it, per the author's own tooltip.

Outputs

text is the string you wire downstream - into a prompt field, a save node, or a second LLM node for a two-stage rewrite. raw_json is the untouched response for debugging. usage_json reports tokens, which is the only way to notice that a long brief is being re-billed on every iteration. credits_consumed and credits_left are floats, also painted on the node after each run.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next

Manager works too - search "KIE.ai Nodes Next". Restart ComfyUI after either route. The dependency list is four small packages (requests, Pillow, numpy, PyYAML), because nothing runs locally. Then set your key once in Settings → KIE.ai Nodes Next → Connection; the pack validates it against KIE before saving, keeps it out of workflow JSON, and only shows the frontend a masked fingerprint. Servers can use KIE_API_KEY instead - it takes priority.

If the node isn't in your menu

This one isn't part of the pack's offline bootstrap. The pack indexes KIE's live API documentation on startup and enriches the catalog from each model's own page, so some nodes only appear once the first deep sync completes. If the Gemini folder is missing entries, restart ComfyUI - or run the catalog refresh node - and check again. After that it re-checks at most once a day, which is how a model KIE adds next month becomes a node without the author shipping a new version.

Worth flagging

• Reasoning controls aren't exposed on the Gemini nodes. The pack gives reasoning-effort and web-search switches to its OpenAI/Codex-style nodes; on Gemini you get prompt, system prompt, history and tools. If you need a reasoning dial, that's a different node in a different folder.
• Token spend compounds. A Pro model in an iterative loop is the single easiest way to burn a credit balance without generating a single image.
• The filter is at the source. This is Google's model behind a reseller; the reseller's looser moderation doesn't change what the model itself will refuse.