Skip to main content

Models

Every agent has a configured model, set in Settings → Model. The choice drives how fast the agent responds, how much it costs over long runs, how deep it can reason, and how much context it can hold.

The picker groups models by provider and shows context window size and per-token pricing where it applies. The exact catalog depends on your account and deployment — which providers are configured, whether subscription-backed or bring-your-own-key models are enabled — so don't be surprised if someone else's list looks different.

Choosing

If you need...Prefer...
Fast, cheap routine workA lightweight or mini-tier model
Good everyday coding/writing balanceA mid-tier general-purpose model
Hard debugging or complex planningA stronger reasoning-oriented model
Very long conversations or large artifactsA model with a larger context window
Predictable cost for frequent scheduled workA cheaper model at lower reasoning effort

Cost deserves a special note for background agents: an agent that runs every 15 minutes multiplies whatever it costs per run. For monitoring and triage, a cheap model usually beats a brilliant one.

And you don't have to pick one perfect default — the usual Taurus pattern is multiple agents with different models: a cheap implementer doing the legwork, a strong reviewer catching its mistakes.

Reasoning Effort

Models that support it get a Reasoning Effort dropdown. Blank uses the model's default; lower is faster and cheaper, higher is slower and more thorough. The labels are provider-native (low / medium / high, sometimes minimal or max).

Treat it as a per-task knob: low for routine edits, high when the problem is genuinely hard enough to pay for the extra thinking.

Pricing

The price in the picker is for comparing models, not predicting your bill — actual run cost depends on input size, output size, tool activity, and run length. Check the run footer and the Billing page for real spend. Subscription-backed models may not show per-token pricing at all.

Model IDs

Models carry provider-prefixed IDs like anthropic/..., openai/..., openrouter/.... You normally just use the picker, but the full prefixed ID is the model's identifier if you ever need to reference one precisely.

Tuning in practice

Start with a sensible model for the agent's role and run a real task. Too slow or expensive → move cheaper. Missing important reasoning → move stronger or raise effort. Constantly running out of context → compact earlier, or pick a larger-context model. Agents are cheap to create, so it's easy to compare empirically.