Model synthesis

The three synthesis ids

Each id is a curated panel + synthesizer pairing. Send it anywhere a model id goes on the OpenAI Chat wire:

The compositions are curated for answer quality per dollar and will evolve as models come and go — the disclosure header below always tells you exactly who participated in your turn. ih-synthesis-max runs frontier models on both stages; picking it is the consent to frontier billing on every participant.

curl https://app.inferencehub.tech/v1/chat/completions \
  -H "Authorization: Bearer sk-prov-live-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ih-synthesis",
    "stream": true,
    "messages": [{ "role": "user", "content": "Design a rate limiter for a multi-tenant API." }]
  }'

Or bring your own panel

The model_synthesis tool gives you the same orchestration with explicit participants: name up to four analysis_models and a synthesizer_model, and set the request's model to the synthesizer. Any model you can reach on the chat wire can sit on the panel.

curl https://app.inferencehub.tech/v1/chat/completions \
  -H "Authorization: Bearer sk-prov-live-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5.2",
    "stream": true,
    "messages": [{ "role": "user", "content": "Design a rate limiter for a multi-tenant API." }],
    "tools": [{
      "type": "model_synthesis",
      "analysis_models": ["glm-5.2", "kimi-k2.6"],
      "synthesizer_model": "glm-5.2"
    }]
  }'

Pricing: each participant at its normal rates

Synthesis has no bundle price and no markup. Every participant — each panel model and the synthesizer — bills as its own line item at that model's normal per-token rates, exactly as if you had called it directly. A synthesis turn therefore costs roughly the sum of asking each panel model yourself, plus the synthesizer pass over the candidates.

• Pay-as-you-go only (for now) — synthesis turns bill your prepaid balance and are not covered by subscription plan windows.
• Line items are linked — the portal's usage view shows one row per participant, tied to the request that spawned them, so you can audit exactly what a synthesis turn cost.
• A failed panelist bills only what it used — a candidate that errors or times out settles its actual tokens (often zero) and the synthesis continues without it.

How a turn runs

• Fresh questions fan out — a new user turn goes to the whole panel in parallel; the synthesizer then streams the merged answer.
• Tool follow-ups don't — when your agent returns tool results mid-conversation, the turn goes to the synthesizer alone (your function tools are its business, not the panel's), so agent loops don't multiply every step's cost.
• Survivors continue — if some panelists fail, synthesis proceeds on the rest; if every panelist fails, the synthesizer answers solo and the response discloses it. You still get an answer, never a wasted turn.

In coding agents

Any coding agent that speaks OpenAI Chat Completions against a custom base URL can use synthesis today — add a synthesis id as a custom model in Cursor or on Kilo Code's OpenAI-wire provider, exactly like any other model. The turn rules above are what make this practical: only your fresh prompts pay the panel's cost and latency, and every tool step in an agent loop runs the synthesizer alone at its normal single-model price and speed.

You always see who served you

Like Auto routing, synthesis is not a black box:

• the x-ih-synthesis response header names every participant, e.g. panel=kimi-k2.6,glm-5.2; synthesizer=glm-5.2;
• panel failures and timeouts are disclosed as SSE comments in the stream (a spec-mandated parser no-op — your client ignores them unless you look);
• the final usage chunk aggregates the whole turn's tokens, and the portal shows the per-participant breakdown.