Explore mode: one prompt, up to five answers

Explore mode sends the same prompt as two to five side-by-side variants, and you pick a winner when they're done or merge them together. It's built for the question you'd otherwise paste into several browser tabs: same prompt, different treatments, answers next to each other instead of scattered across windows that don't know each other exist.

The setup is the "Explore this prompt" dialog. The Compare models preset comes pre-filled with up to three models from different providers, so the common case, "ask three strong models and compare", is ready without configuration. Each variant row has its own model picker; chat models are listed first, and coding agents like Cursor are labeled separately as agents, because they behave differently. If a variant can't actually run, for example Cursor without a GitHub repository in your prompt, the dialog tells you before anything launches instead of failing mid-comparison. Variants run without a persona by default so you're comparing the models themselves, not one persona's instructions repeated three times; you can still pick a persona per row when that's the point.

Then the variants stream in as columns. If one fails, its column says so plainly, with a retry, and the failed branch is left out of scoring and merging automatically.

Scoring, picking, or merging

When every variant finishes, GPTree scores the answers automatically: a winner badge, per-variant scores, and a written rationale explaining the call. The scoring reads the full answers against your original prompt, so a long answer that actually covers the task outranks a short one that only sounds confident. Failed variants score zero with the failure as the reason. From there you can accept the winner and continue in that branch, or ignore the verdict and press "Use this answer" on the column you actually believe.

Sometimes no single column deserves to win, and that's what merging is for. "Merge all into one answer" appears once everything settles: it collects every successful answer and opens the merge dialog in one step. You can also curate by hand, adding whole responses, highlighted selections, or the compare drawer's key-difference summaries to My Answer. Summary picks carry a size chip, for example "888 of 11,333 chars", with a one-click "Use full text" swap, so a compressed summary never quietly stands in for the full answer.

The merge dialog asks one question worth answering: what should the final answer be? Tell it "merged feedback to send back to the spec's author" and the result is written as that deliverable, drawing on your original prompt and every source. It runs on the best model among your sources by default, sizes its length to the material instead of a fixed cap, and shows which picks contributed. Copy the result as Markdown right from the preview, or save it into your tree.

Where Explore stops

Five variants is the ceiling and two is the floor. If you want fifteen takes, that's several runs.

Scoring waits for your slowest model, since a verdict on half the answers wouldn't mean much. Failed variants don't hold it up; they're excluded and the rest are scored.

And mind the credits. Every variant is its own model run, so a five-variant Explore is five model runs, and premium models use more credits. Usage shows up in your dashboard. Explore is best saved for prompts where the answer matters enough to buy it five times.

One more boundary: Explore answers one prompt many ways. The moment you want to take an answer somewhere new, you're back in branching territory, and that's by design. Branching splits an existing conversation into diverging paths you walk one at a time, while Explore multiplies a single prompt into parallel, side-by-side takes. If two branches have already grown apart and you're weighing them against each other, that's the two-way compare, which has its own article.