The dial

Every agent runs on a model. Today you pin one; with the dial you say how much a miss would cost you and connect0 picks the cheapest model that is strong enough for the task. Nothing changes for an agent until you turn it on.

Four positions

The dial has four positions. Pick by how much a wrong answer costs you, not by how big the task looks:

Position	When to use it
Low	You know exactly what you want. Well-defined, small, easy to check.
Medium	You know roughly what you want. Messy or multi-part; the model fills in steps.
High	A subtle miss is expensive. More reasoning, closer to review-ready.
Ultra	The outcome is clear, the path is not. Open-ended, cross-cutting.

Start on medium. Turn it down when the task is clear, up when a miss costs more than the wait. Higher positions do not use "more of the same" model — each position sets a minimum strength the chosen model must reach for the agent's kind of work.

Task classes

Models are strong at different things. The dial ranks candidates for the agent's task class — conversation, classification, extraction, summarisation, writing, reasoning, coding, tool use, long context, or vision — using strength profiles the platform keeps per model and class. Set the class on the agent's model tab, or choose Auto and a small classifier decides per turn (below).

What the dial does on each run

1. Works out the position: a run-level override wins, then the thread's position (when the agent allows it), then the agent's own.
2. Applies your pins, if any: a pin on the agent wins over a pin on the account, which wins over the platform's default.
3. Otherwise ranks every model your account may use: drops any that cannot handle the turn (context size, tools, images), keeps those at or above the position's strength floor, and picks the cheapest on your plan — platform price on credits, or the fee plus the provider's bill on your own key. If nothing reaches the floor it relaxes the floor by one band and says so; if still nothing fits it falls back to the agent's pinned model.
4. Keeps the next two ranked models as fallbacks, so a provider error walks to the next-cheapest admissible model rather than a hand-typed list.

Every routed run records the decision. Ask Zero to explain a run, or open the run's audit trail, and you will see a line such as Routed on the dial (medium, coding): openai/gpt-5-mini over anthropic/claude-sonnet-5 — same band, 40 % cheaper on your plan.

Turning it on

• Agent — the model tab has a control Pinned model | Dial. Under Dial: the position, the task class, and Advanced → pins per position when you want one model at, say, high. The pinned model stays as the fallback.
• Account — Settings → Models → Dial sets pins per position for every agent on the dial that has no pin of its own, and Optimise for: total cost, or only what the platform bills in credits.
• Per run — the API's run body and the agent's MCP tool accept a dial value (low | medium | high | ultra) for one run.
• In a chat — when the agent allows it, the composer's mode chip sets the thread's position.

Auto — classify each turn

With the task class on Auto, a small classifier looks at the turn before the main call: what kind of work it is and how hard. It runs only on turns long enough to matter, is cached for a day per agent and prompt, never spends more than a few hundred tokens on a small model, and can move the dial one step from the agent's position when it is confident. If it is unsure or unavailable, the run uses the agent's class as before and the trace says the classifier stood aside.

Costs

The dial is designed to spend less, not more: it prefers the cheapest model that is strong enough, and the trace estimates what the same run would have cost on the dearer candidate. The classifier's own call is metered like any other model call and marked as the classifier in usage.