> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://openrouter.ai/docs/llms.txt. > For full documentation content, see https://openrouter.ai/docs/llms-full.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://openrouter.ai/docs/_mcp/server. # Fusion Server tools are currently in beta. The API and behavior may change. The `openrouter:fusion` server tool gives any model access to multi-model deliberation. When your model decides a prompt benefits from multiple perspectives, it invokes this tool — a panel of models answers in parallel, a judge compares their responses, and the structured analysis comes back to your model for the final answer. This is the same pipeline behind the [`openrouter/fusion` model alias](/docs/guides/routing/routers/fusion-router) and the [`fusion` plugin](/docs/guides/features/plugins/fusion). Using the server tool directly gives you the most control: choose your own outer model, combine it with other tools, and configure the panel and judge independently. ## Quick start ```typescript title="TypeScript" const response = await fetch('https://openrouter.ai/api/v1/chat/completions', { method: 'POST', headers: { Authorization: 'Bearer {{API_KEY_REF}}', 'Content-Type': 'application/json', }, body: JSON.stringify({ model: '{{MODEL}}', messages: [ { role: 'user', content: 'Survey the strongest arguments for and against a carbon tax. Where do experts disagree?', }, ], tools: [ { type: 'openrouter:fusion' }, ], }), }); const data = await response.json(); console.log(data.choices[0].message.content); ``` ```python title="Python" import requests response = requests.post( "https://openrouter.ai/api/v1/chat/completions", headers={ "Authorization": f"Bearer {{API_KEY_REF}}", "Content-Type": "application/json", }, json={ "model": "{{MODEL}}", "messages": [ { "role": "user", "content": "Survey the strongest arguments for and against a carbon tax. Where do experts disagree?", }, ], "tools": [ {"type": "openrouter:fusion"}, ], }, ) print(response.json()["choices"][0]["message"]["content"]) ``` ## When does the model invoke it? The tool's description tells the model to call `openrouter:fusion` only when a task genuinely benefits from multiple perspectives — research questions, multi-domain critique, "compare and contrast" prompts, or anything where being wrong is expensive. Simple tactical prompts won't trigger it. To **force** fusion on every request, set `tool_choice: "required"`. See [Forcing fusion on every request](/docs/guides/routing/routers/fusion-router#forcing-fusion-on-every-request). ## Parameters Pass an optional `parameters` object on the tool entry to override defaults: ```json { "tools": [ { "type": "openrouter:fusion", "parameters": { "analysis_models": [ "~google/gemini-flash-latest", "deepseek/deepseek-v3.2", "~moonshotai/kimi-latest" ], "model": "~anthropic/claude-opus-latest" } } ] } ``` | Field | Default | Description | | ----------------------- | --------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `analysis_models` | Quality preset (`~anthropic/claude-opus-latest`, `~openai/gpt-latest`, `~google/gemini-pro-latest`) | Models that form the panel. Each runs in parallel with `openrouter:web_search` and `openrouter:web_fetch` enabled. 1–8 models allowed. | | `model` | Your outer model | The judge that produces the structured analysis JSON. Defaults to the same model handling your request. | | `max_tool_calls` | `8` | Max tool-calling steps each panel model and the judge may take in their `openrouter:web_search` / `openrouter:web_fetch` loop before they must return text. Range 1–16. | | `max_completion_tokens` | Provider default | Max output tokens (including reasoning) per inner panel/judge call. Keeps reasoning-heavy models from exhausting their budget before producing visible text. | | `reasoning` | Provider default | Reasoning config forwarded to the panel and judge calls — an object with optional `effort` and `max_tokens`. | | `temperature` | Provider default | Sampling temperature (`0`–`2`) forwarded to the panel and judge calls. | ## What the tool returns On success, the tool result contains the structured analysis and the raw panel responses: ```json { "status": "ok", "analysis": { "consensus": ["Points all or most panel models agreed on"], "contradictions": [ { "topic": "...", "stances": [{ "model": "...", "stance": "..." }] } ], "partial_coverage": [ { "models": ["..."], "point": "Only some models covered this" } ], "unique_insights": [ { "model": "...", "insight": "Something only one model raised" } ], "blind_spots": ["Topics no panel model addressed"] }, "responses": [ { "model": "anthropic/claude-opus-4.5", "content": "..." }, { "model": "openai/gpt-4.1", "content": "..." }, { "model": "google/gemini-2.5-pro", "content": "..." } ] } ``` When some panel models error but at least one succeeds, the result still has `status: "ok"` and adds a `failed_models` array describing which ones failed and why. ### Judge degradation If the **panel** succeeds but the **judge** fails — an upstream error, an empty completion, or output that isn't valid analysis JSON — the tool does **not** error. It returns `status: "ok"` with the raw panel `responses` and simply **omits** `analysis`. Your model can still write the final answer from the panel responses: ```json { "status": "ok", "responses": [ { "model": "anthropic/claude-opus-4.5", "content": "..." } ] } ``` ### Hard failures The tool only returns `status: "error"` when it can't produce any useful output. In that case it includes a typed `failure_reason`: ```json { "status": "error", "error": "all panel models failed", "failure_reason": "all_panels_failed" } ``` | Reason | Meaning | | -------------------------- | ------------------------------------------------------------------------------- | | `all_panels_failed` | Every panel model returned an error. | | `insufficient_credits` | Every panel model failed and at least one was due to insufficient credits. | | `rate_limited` | Every panel model failed and at least one was rate-limited. | | `fusion_invocation_capped` | Fusion was already invoked earlier in the same turn; a second call is rejected. | | `unexpected_error` | An unexpected error interrupted the fusion run. | The calling model can fall back to answering without the analysis whenever fusion fails or degrades. ## Web tools `openrouter:web_search` and `openrouter:web_fetch` are enabled on both the **panel** and the **judge** calls, so models can pull fresh sources while they answer and analyze. The judge compares the panel responses rather than merging them: it treats what all or most models agree on as higher-confidence consensus, surfaces contradictions, preserves unique insights from individual models, and flags blind spots none of them addressed. The outer model writes the final answer from that analysis — so the result isn't a simple majority vote. ## Recursion protection Inner fusion calls carry an `x-openrouter-fusion-depth` header. Panel and judge models cannot recursively invoke `openrouter:fusion` — the plugin refuses to inject the tool a second time, keeping deliberation bounded to a single level. ## Related * [Fusion Router (`openrouter/fusion`)](/docs/guides/routing/routers/fusion-router) * [Fusion plugin](/docs/guides/features/plugins/fusion) * [Web Search server tool](/docs/guides/features/server-tools/web-search) * [Web Fetch server tool](/docs/guides/features/server-tools/web-fetch) * [`/labs/fusion`](/labs/fusion) — interactive playground