Turbo: better answers on hard questions

Turbo spends extra compute on a single turn: it generates several answers, has a judge pick the strongest, and merges the best parts into one reply. This page shows you where to configure it, what each setting actually does, and the one situation where you must switch it off.

Where is it?

The lightning chip in the composer. It reads Turbo Auto by default. Click it to open the settings popover.

Mode has three values:

What does Auto decide on?

Auto scores each prompt locally before the turn runs — no extra model call, no cost. The signals it looks at are:

• Prompt length (a long prompt scores higher than a short one)
• Code, science or maths hints in the text
• Comparison or decision wording — “versus”, “trade-off”, “which is better”
• Number of stated requirements — bullets, numbered lines, “must”, “ensure”
• Number of questions in the message
• How large the conversation already is

Score high enough and Turbo fires. In practice: a one-line question stays fast and cheap, a multi-part design question gets the extra passes. You do not have to decide.

The three settings

Intensity — how many candidate answers get generated:

Diversity — how different the candidates are from each other, as a temperature spread:

Strategy — how the winner is chosen. Two options are selectable: Automatic and Judge (general). Both currently resolve to the same judge, which scores the candidates and picks the strongest, so the choice does not change the result today.

Enhanced final synthesis is a switch, on by default. With it on, a final pass merges the strongest parts of the candidates into one answer instead of returning only the winning draft. Leave it on unless you specifically want to see a single unmodified candidate.

The trade-off: Turbo turns have no tools

When Turbo fires, that turn runs without tools. No web research, no memory tools, no quantum backend listing. If you need the assistant to look something up, recall a stored fact, or list quantum providers, set the lightning chip to Off for that turn.

This is the single most important thing on this page. A question that begins “search for…” or “what backends are available…” should be asked with Turbo Off.

What does it cost?

More candidates means more generation time and more of your message quota consumed on that one turn. Max (5 candidates) plus synthesis is the slowest and most expensive combination; Balanced (3) is the default because it is the reasonable middle.

Your quota counters are described in Chat limits and errors.

Worked pairing

Use Turbo for this:

Compare an event-sourced design against a CRUD design for an audit-heavy
billing service. Cover write throughput, replay cost, and how each one
handles a schema change. Recommend one and say what would change your mind.

Set Intensity to Balanced, leave Enhanced final synthesis on. This is exactly the shape Auto already fires on: comparison wording, several requirements, no lookup needed.

Switch Turbo Off for this:

Which quantum backends are currently available through QCOS?

That question needs a tool call, and a Turbo turn has none. See Quantum backends.

Related

• Choosing a model — Oracle plus Turbo for the hardest reasoning turns.
• Memory and personalization — memory tools are among the tools a Turbo turn gives up.