Compare answers from multiple models, then continue with the preferred conversation

[ElliotDrel]

Note

?? gpt-6.1-sol responding on behalf of Elliot

Problem

I want to ask two or more models the same question, compare their answers, then continue the conversation with the model whose answer I prefer. Manually creating threads and copying the prompt/context makes this cumbersome.

Proposed behavior

1. Choose ?Compare models? in the composer and select two or more available provider/model configurations.
2. Submit one prompt with the same attachments and explicitly shared context to independent runs. Start with new conversations; branching an existing conversation can follow once cross-provider context semantics are defined.
3. Show labeled, independently streaming answers side by side on wide screens and in switchable panels on mobile. Show each run's completion, failure, cancellation and available usage information.
4. Offer ?Continue with this answer? on each result. Open that model's actual conversation, preserving its context and tool history. Subsequent messages go only to the selected conversation.
5. Keep other answers accessible; let users cancel remaining runs without deleting their results. One failed model must not discard successful answers.

Smallest useful scope

Start with two models answering questions/planning in a mode that prevents workspace mutations. The user chooses the preferred answer; automatic judging and merging answers are optional future features. Show that multiple runs consume each provider's allowance independently.

For a later coding mode, each candidate needs an isolated checkout from the same baseline; selection of a conversation and application of its code changes must be separate explicit actions. Concurrent candidates must not edit the same working directory.

Support web/desktop and mobile, including remote connections, through server-owned runs so a client disconnect does not lose results. Only offer configurations actually available in the selected environment.

Cursor research

• Cursor's explanation of its original parallel-agent flow: identical prompts run independently, each coding agent uses an isolated Git worktree, and users review and apply the preferred result.
• Cursor's April 2026 clarification: the former Multiple models picker was removed; the newer best-of-n flow differs from manual side-by-side comparison. Separate local/cloud runs are the described alternative.

These document observable product behavior, not Cursor's private implementation. T3 should preserve the explicit human choice and continuation workflow requested here.

Related request

#6911: parallel agents within a thread focuses on different agents tackling separate subtasks and returning results to a shared thread. This proposal instead sends the same question to independent models and lets the user choose one conversation to continue.

[ElliotDrel]

Note

🤖 gpt-6.1-sol responding on behalf of Elliot

Would maintainers approve an initial two-model, new-conversation comparison restricted to provider modes that enforce non-mutating work? Reuse ordinary durable threads and select a result by opening its actual thread; no automatic judge, answer merge, code application or existing-thread branching. Each run must retain independent cancellation/failure and survive client disconnect. If that direction fits, the remaining scope decision is whether mobile gets a result switcher in the first increment or the action stays unavailable there until supported.