Changing the model, end to end

What you are trading#

A realtime model is not a component that can be chosen on price alone. Each one refuses a different set of things, and the refusals arrive as error codes on live calls rather than as a feature matrix. The two the Voice Lab offers differ like this.

Before you switch#

1. Check the model id itself. A Vertex-style path or a malformed id is refused as bad_model before any call row is created — the fifth Lab call rang the owner and then died with a 1008 because the id was a Vertex path.
2. Check what the line depends on. If the supervisor panel is part of how that line is run, an immutable session takes it away.
3. Check the budget. Live calls are billed on audio tokens — roughly 4× the same model's text rate on input and 2.7× on output — so a model with a lower text rate is not automatically cheaper on the phone.
4. Note today's numbers: reply median, the first-token share of the wait, and cost per call from the profiler. A switch with no baseline cannot be judged, only defended.

The chain a switch runs through#

1. TRIGGER — the model is changed in the Voice Lab, at the workspace or on one profile.
2. USER / EXTERNAL EVENT — a person makes the change; the first affected caller is an ordinary inbound or outbound call.
3. AUTH / WORKSPACE RESOLUTION — the setting is written inside the owning workspace; both audiences reach the Lab through /api/voice/realtime/lab*.
4. INGEST / REQUEST — the next call's ring validates the model id and refuses bad_model before a Call row exists, so a malformed id costs nothing.
5. CANONICAL RECORD — the call is created and the resolved settings, the model among them, are snapshotted on call.quality.
6. CLASSIFICATION / RESEARCH / REASONING — the model now does the listening and the answering; the worker holds no business logic of its own.
7. KNOWLEDGE + MEMORY + RULES — the same brief is rendered as before. Only the model reading it has changed.
8. AUTONOMY / APPROVAL — unchanged. Outbound still runs through comms_providers.dispatch_call, so memory blocks, consent, autonomy and budget gate the call exactly as before.
9. ACTION / PROVIDER — the session opens against the new model. One that refuses to start is opened once more on the same model and once on the fallback before the call is given up.
10. RESULT — model_session_restarted goes on the record when a restart happened, so a call that sounded fine but restarted twice is visible afterwards.
11. RELATIONSHIP / TIMELINE / MEMORY — summary, lead, memory and follow-up run through the same voice.end_call whichever model held the call.
12. AUDIT / USAGE / NEEDS YOU — model usage is booked into costs split by modality; sessions the model refused escalate into Needs You as engine_error.
13. NEXT — compare reply median and cost per call against the baseline, over calls rather than over one call.

Afterwards: what proves it#

Three numbers, all of them already recorded. Reply median comes from the turn_timing events rather than from transcripts. Cost comes from GET /api/usage/voice, the per-call profiler, which reports what was measured and says what was not rather than estimating the gap — that matters because the SDK reports input_audio_tokens and session_duration but no input_audio_duration, and code that looked for one recorded 0.0 seconds for every call ever metered.

The third is the share of the wait that belongs to the model. When the first token is 60% of it, the review says so and offers no control to turn; that share moving is the clearest single signal that a model change was the right one.

What can fail#

Calls die shortly after ringing

The model id is wrong for this stack: refused as bad_model where the check runs, a 1008 on the wire where it does not.

The supervisor panel stopped working

The new model holds an immutable session. Guidance is recorded as guidance_undeliverable rather than silently dropped.

Cost rose while replies got shorter

Audio tokens, not words. A Live session is re-billed for its whole context every turn, so a long conversation costs more per reply than a short one on the same model.

A call reports no tokens

It is still charged from its own duration at the published per-minute rate. Zero is not an outcome the ledger accepts.

The first call after the switch was slow

Greeting warm-up is per process, and a newly started process has an empty LineCache until prewarm fills it.

Questions#

Related#

• What a setting cannot change
• The realtime voice engine
• What a call costs
• The voice model refused the session
• Findings a setting cannot fix
• Changing what the agent is told
• Changing a provider, end to end