Roundtable

POST /v3/roundtable sends one prompt to several models in parallel and brings their answers back together. Instead of trusting a single model, you get a panel: ask three models the same question and either merge their responses into one consolidated answer or compare them side by side. It runs on the same gateway, key, routing, and tracing as every other call. Use it when an answer is high-stakes or subjective and a second (and third) opinion is worth the extra cost: architecture decisions, evals and grading, content review, or any “let’s see what a few good models think” moment.

​

How it works

1. Fan out — your input goes to every model in models, in parallel.
2. Resolve — depending on the resolution mode, the results are consolidated into one answer, returned as-is, or voted into a single choice.
3. Return — you get the consolidated data, aggregated meta (models used, total cost, duration), and the per-model model_results.

​

Resolution modes

In summary mode you can override the consolidation model with summary_model. In multiple_choice mode, pass the literal options in choices.

​

Shared options

These apply across every model in the panel: Each entry in models can also override temperature, max_tokens, top_p, and reasoning effort for that model alone.

​

What’s next