Model

The Model page picks the language model that drives the agent. 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. Everything else (the voice, the tools, the knowledge base) sits around it. Choosing the model is choosing how fast, how reliable and how expensive each call is.

What the Model page does

The page shows a picker of the models available to your workspace. Each entry lists three things:

• The model name (for example, Claude Haiku 4.5).
• The tier that sets its price band: Fast, Balanced, Power or Speed.
• The per-minute price, shown next to the model so you can compare cost at a glance before you pick.

You select one model. That choice lands in a draft and changes nothing on live calls until you publish.

The tiers

Every model belongs to a price tier. The tier is derived from the model's per-minute voice rate, so a tier label always means the same price band no matter which provider the model comes from.

Available models

Two providers back the models, and they differ on where the call's data is processed. Claude models run on AWS Bedrock in Sydney, so inference stays in Australia (the au. region prefix). Gemini models run on Google Vertex; the Fast Gemini Flash tier is hosted in-region, while the larger Gemini tiers may process outside Australia.

The recommended default: Claude Haiku 4.5

New agents start on Claude Haiku 4.5, the Fast tier on AWS Bedrock. It is the right starting point for almost every agent, and here is why.

On a phone call the model's most important job is to call the right tool at the right moment: looking up an order, booking a meeting, transferring to a human. A model that answers beautifully but reaches for the wrong tool (or no tool) gives the caller a wrong answer. In our testing Claude is the most reliable model at tool-calling, and Haiku is both fast and dependable at it. The Gemini tiers are offered and can be cheaper or quicker, but we find them less reliable at tool calls, so reach for one only when you have a reason.

The second reason is latency. The caller hears every pause. A fast, reliable answer beats a slower, marginally smarter one, because the lag is something the caller feels on the line and the extra cleverness usually isn't. Haiku starts speaking quickly, which is why it is the default rather than a heavier model.

The trade-off: latency, quality and cost

Choosing a model is balancing three things.

• Latency is how fast the model starts responding. On a call this is the difference between a natural exchange and an awkward gap. Faster models feel more human. Latency is part of call quality, not separate from it.
• Quality is how well the model follows the prompt, picks the right tool and recovers when a call goes sideways. More capable tiers have more reasoning headroom for messy or unusual calls.
• Cost is the per-minute price, shown next to each model in the picker. A Power model can cost several times what a Fast model does per minute.

A heavier model is not automatically better on a call. The extra reasoning often buys little for a routine enquiry while adding latency the caller can hear and cost you pay every minute. Start fast, and move up only when the agent is genuinely getting calls wrong.

Pricing

Voice is billed per minute of audio, and the rate is set by the model's tier. One credit equals one Australian cent.

The price you see next to each model in the picker is the exact rate the meter bills, read from the same place, so the displayed price can never drift from the invoice. Token prices for the underlying models are synced from OpenRouter once a day; they are not fetched live on each call, so the per-minute price you pick is stable for the call. For how minutes turn into credits and how to read your usage, see Billing and credits.

When to change the model

Change the model when you have evidence, not a hunch. The symptom points at the fix:

• The agent picks the wrong tool, misreads the caller, 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 the agent handles only simple, repetitive calls: a Speed-tier model may be enough.

Don't judge a model swap off a single call. Use the test panel to talk to the draft, and run evals to compare the old model against the new one across the same scenarios. A model is reliable only if it behaves correctly on every call, not most of them, so trust a repeated run over one good or bad result.

Drafts and publishing

Selecting a model edits a draft of the agent. Live calls keep using the currently published model until you publish the draft. Publishing makes the new model take over the next call, so test and run your evals first, then publish. For how drafts and publishing work across the whole agent, see Agents.

Related

• Agents: drafts, testing and publishing.
• Tools: what the model calls; tool-calling reliability is the main reason Haiku is the default.
• Evals: compare models with evidence before you switch.
• Billing and credits: how per-minute pricing turns into credits.