Model
Picking the language model that powers your agent, and the trade-off between latency, quality and cost.
The model is the brain of the call. It reads the system prompt and the conversation so far, decides what to say, and decides which tool to call.
What the picker shows
The Model page lists the models available to your workspace. Each entry shows:
- the model name, for example Claude Haiku 4.5;
- the tier: Fast, Balanced, Power or Speed;
- the residency, Australia or global;
- the per-minute model rate.
You select one model. That choice lands in a draft and changes nothing on live calls until you publish.
Every model, with its tier, residency and rate, is on Available models.
The tiers
| Tier | What it is for |
|---|---|
| Fast | The everyday default. Quick to respond, dependable on tool calls, low cost. |
| Balanced | More reasoning headroom at a higher rate, still quick enough for a call. |
| Power | The most capable models for the hardest calls, highest rate, usually slowest. |
| Speed | The lightest, lowest-latency models for simple, high-volume calls. |
New agents start on Claude Haiku 4.5 (Fast). In our testing Claude is the most reliable at tool-calling, and Haiku is fast enough that the caller does not hear a gap. Start here, and move up only when the agent is genuinely getting calls wrong.
When to change the model
Change the model when you have evidence, not a hunch:
- The agent picks the wrong tool or fumbles an unusual call: try a stronger tier.
- Replies are correct but the caller is left waiting: try a faster model.
- Cost is the concern and calls are simple and repetitive: a Speed-tier model may be enough.
Use the test panel and evals to compare the old model against the new one. A model is reliable only if it behaves correctly on every call, not most of them.
What a minute costs
The per-minute rate shown in the picker is the model rate only. Your invoice also carries transcription and speech synthesis charges. The full breakdown is on Billing.