FastRouter Blend: Multi-Model Deliberation

Multi-model deliberation — ask a panel of models the same prompt in parallel, then have a judge model compare their answers into a structured analysis.

Overview

Blend asks a panel of models the same prompt in parallel, then a judge model compares their answers into a structured analysis: agreements, disagreements, coverage gaps, standout insights, missing considerations, and confidence notes.

Instead of trusting a single model, Blend gathers multiple independent perspectives on your prompt and has a judge model analyze how they compare. The judge does not merge the answers — it evaluates them, so you can see where models agree, where they conflict, and what each one uniquely contributed.

• Parallel panel calls — the same prompt runs against every panel model at once.

• Structured judge analysis — a single JSON object comparing the answers.

• Aggregated usage & cost — panel + judge tokens and spend, summed and itemized.

• Reuses routing, BYOK & credits — sub-calls run through the normal gateway path.

Two ways to use Blend

Method	How to trigger	When to use
Model alias — fastrouter/blend	Set "model": "fastrouter/blend".	Blend always runs and returns a human-readable summary as the completion content. Best when you want Blend to be the whole response.
Server tool — fastrouter:blend	Attach a fastrouter:blend tool to a normal request.	The outer model decides when to call it; the structured analysis is fed back as a tool result so the model writes the final answer.

---

1. Model alias — fastrouter/blend

Send a normal chat completion request with the model set to fastrouter/blend:

curl https://api.fastrouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "fastrouter/blend",
    "messages": [
      { "role": "user", "content": "Survey the strongest arguments for and against EV cars." }
    ],
    "analysis_models": ["anthropic/claude-sonnet-5", "openai/gpt-5.2", "google/gemini-3-pro"]
  }'

The assistant content is a markdown summary: a ## Panel responses section (one subsection per model) followed by ## Analysis containing the structured JSON. analysis_models is optional — see Panel & judge selection.

2. Server tool — fastrouter:blend

Attach the hosted fastrouter:blend tool to a request that targets any normal model. Blend parameters go inside the tool's parameters object.

curl https://api.fastrouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "x-ai/grok-4.20-beta",
    "messages": [
      { "role": "user", "content": "Assess the case for and against banning internal combustion engine vehicles by 2035." }
    ],
    "tools": [
      {
        "type": "fastrouter:blend",
        "parameters": {
          "analysis_models": [
            "google/gemini-3-flash-preview",
            "z-ai/glm-5.2",
            "moonshotai/kimi-k2.7-code"
          ],
          "model": "anthropic/claude-sonnet-5"
        }
      }
    ]
  }'

What happens under the hood:

1. The gateway replaces the hosted tool with a callable function named fastrouter_blend (providers reject unknown hosted tool types).

2. The outer model runs normally and decides whether to call fastrouter_blend.

3. When it does, the gateway runs the Blend pipeline, injects the structured BlendResult JSON as the tool result, and re-calls the model.

4. The model uses that analysis to write the final answer.

Model-driven. With the server tool, Blend runs only if the outer model chooses to call fastrouter_blend. For simple prompts a model may answer directly and skip it. Use the model alias fastrouter/blend at the top level if you want Blend to always run.

---

Parameters

For the alias these go at the top level of the request body. For the tool they go inside the tool's parameters object.

Parameter	Type	Required	Description
analysis_models	string[]	No	The panel models that answer in parallel. Capped at 5. If omitted, the panel is auto-selected from the router's top candidates (capped at 3).
model	string	No	The judge model that compares the panel answers. Defaults to the top-ranked panel model when omitted.

Everything else in your request — the messages, plus any other tools, sampling params, or response_format — is forwarded to each panel model. Only the model id is swapped, the Blend tool is stripped, and streaming is turned off for the sub-calls.

Panel & judge selection

Panel models

• Supplied: your analysis_models are used as-is, in order, de-duplicated, capped at MAX_BLEND_ANALYSIS_MODELS = 5. Supplying 2 uses exactly 2.

• Auto: if you supply none, the panel is derived from the router's classifier top candidates, capped at MAX_BLEND_AUTO_MODELS = 3.

Judge model

• The model override when provided.

• Otherwise, the top-ranked panel model.

---

Response format

Model alias — human-readable content

The completion content is markdown, structured like:

## Panel responses

### anthropic/claude-sonnet-5
<that model's full answer>

### openai/gpt-5.2
<that model's full answer>

**Tool calls**    ← only shown if a panel answered with tool calls
[ { "id": "call_1", "type": "function", "function": { ... } } ]

## Analysis
<the structured judge JSON>

Server tool — structured tool result

The Blend pipeline returns a BlendResult JSON object as the tool result. The model then writes the final natural-language answer from it. The object looks like:

{
  "status": "ok",
  "analysis": { /* judge JSON, see below */ },
  "responses": [
    {
      "model": "google/gemini-3-flash-preview",
      "content": "…",
      "tool_calls": [ /* present only if the panel emitted them */ ],
      "usage": { "prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0, "cost": 0 }
    }
  ],
  "failed_models": [ { "model": "…", "error": "…" } ],
  "panel_models": ["…"],
  "judge_model": "anthropic/claude-sonnet-5",
  "trace_id": "blend-…",
  "usage": { "prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0, "cost": 0 },
  "judge_usage": { "prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0, "cost": 0 }
}

Field	Description
status	"ok" or "error".
analysis	The judge's structured comparison JSON. Omitted if the judge degraded.
responses	The successful panel answers (content and/or tool_calls).
failed_models	Panel models that errored, with the error message.
panel_models / judge_model	Which models were used (observability).
trace_id	Shared id correlating every panel + judge sub-call in the run.
usage / judge_usage	See Usage & cost.

Judge output (the analysis JSON)

The judge returns only this object — it analyzes, it does not merge:

{
  "agreements": ["points where most or all responses aligned"],
  "disagreements": [
    {
      "topic": "the point of disagreement",
      "positions": [
        { "model": "model identifier", "position": "that model's stance" }
      ]
    }
  ],
  "coverage_gaps": [
    {
      "point": "a relevant point only some responses addressed",
      "covered_by": ["model identifiers that covered it"]
    }
  ],
  "standout_insights": [
    { "model": "model identifier", "insight": "something notable only this model raised" }
  ],
  "missing_considerations": ["important aspects no response addressed"],
  "confidence_notes": ["where responses seemed uncertain, hedged, or speculative"]
}

---

Usage & cost

The usage field is the sum of every panel call plus the judge call — prompt_tokens, completion_tokens, total_tokens, and cost. The judge_usage field isolates just the judge call (its numbers are already included in the total).

Panel usage is counted even for panels that failed to produce usable content, because the tokens were still spent.

Errors & graceful degradation

• Judge fails / returns non-JSON but at least one panel succeeded → status stays "ok", analysis is omitted, and the panel responses are still returned.

• No panel produced output → status: "error" with a failure_reason.

failure_reason	Meaning
all_panels_failed	Every panel model errored.
blend_invocation_capped	Blend was already invoked earlier in this turn (one run per turn).
rate_limited	A rate limit was hit.
unexpected_error	An unexpected internal error (e.g. no models available).

---

Limits & notes

• Max panel size: 5 supplied models, 3 auto-selected.

• Blend runs at most once per turn (recursion protection).

• The server-tool path is non-streaming and top-level only.

• Panel models receive your other tools too, so a panel may answer with tool_calls; those are captured and included in what the judge compares.

• The fastrouter:blend server tool is model-driven; use fastrouter/blend when you want Blend to always run.