Autopilot: AI Crucible Now Plans Every Run for You

Until now, a run on AI Crucible started with choices: which models, which strategy, how many rounds, and who judges. Those choices matter, but most people should not have to make them. Autopilot makes them for you.

You type a question. Autopilot reads it, plans the run, and shows you the plan and its cost before anything happens. You set two things only: speed and cost.

Time to read: 4-5 minutes.

Autopilot plans a run from speed and cost settings

What does Autopilot decide?

Autopilot reads each question for what kind of answer it needs, how hard it is, and what is at stake. It also checks whether the question needs the web, careful reasoning, your files or your connected tools.

From that read, it picks:

The route: one model, or a panel of models from different vendors.
The method: one of the seven ensemble strategies, such as Debate for a decision or Red Team for a critique.
The models: a panel drawn from the quality band your cost setting allows, spread across vendors.
The judge: the model that writes the final answer. It is not a lightweight model unless those are the only models open to you, and then the plan says so.
The rounds, reasoning effort and web search for the run.

For high-stakes questions in medicine, law, finance or construction, Autopilot calls that field's specialist team.

What do the speed and cost settings do?

Speed sets how many rounds a panel may take. Fast allows one round. Balanced allows up to two and Thorough up to three. Hierarchical always takes at least three, because it plans, delegates and reviews. A panel stops early once its members agree, except at Thorough with Max quality.

Cost sets the panel size and the models behind it. Lean uses the smallest panels on fast, low-cost models. Standard uses frontier models. Max quality uses the strongest models, with a larger panel on hard questions.

Questions Autopilot reads as simple and low-stakes go to one model at any setting. At Fast with Lean cost, anything that is not high-stakes goes to one model.

How do I see the plan before it runs?

As you type, a line under the box summarizes the plan. It reads like "Debate · 3 models · ~45s · ~9.0K credits".

Click it to open the plan card. The card shows what Autopilot read in your question, who sits on the panel and what each model does. It also compares the cost with a Max-quality run of the same shape.

From the card you can Ask with this plan, or Open in Advanced to change any part of it by hand.

If your balance cannot cover a plan, Autopilot scales it down to fit and says so on the plan. If Autopilot is unsure how to read a question at Standard cost, it may use Max-quality models, and the estimate shows that too.

What can I do with the answer?

Autopilot puts the answer first. A strip above it shows how the answer was made and what Autopilot saved against Max quality. The strip also holds four actions:

Upgrade reruns the question one cost level higher, or at Thorough once you are at Max quality. It is hidden when the plan was scaled down to fit your balance, or already ran at Max quality and Thorough.
Challenge this answer runs a Red Team panel against it, looking for weak claims and what would change the conclusion.
Second opinion asks the same question of models from vendors the first run did not use.
Show the work opens every model's answer, round by round.

If the plan looked wrong, choose Wrong plan? and say what would have fitted. Those reports help tune how Autopilot plans.

Can I still choose the method myself?

Yes, in three ways. Start a question with a command to force a method: /single, /debate, /redteam, /panel, /collab, /refine, /cot or /hierarchical. Use the team menu to pin a specialist team instead of "Any team".

Or switch to Advanced mode, the full workbench, where you pick every model, strategy, role and round yourself. Nothing about Advanced mode has been removed.

Follow-up questions in the same chat keep the plan when the kind of question stays the same. If you ask something of a different kind, or change a setting, Autopilot plans again.

Can I use Autopilot from the API?

Yes. Send crucible/auto as the model to the chat completions or responses endpoint. Set speed and cost with ai_crucible.speed and ai_crucible.cost, or the x-crucible-speed and x-crucible-cost headers.

The response includes the plan Autopilot ran in ai_crucible.plan. See the developer page for the full specification.