Short answer: A voice agent's latency, cost, and quality trade-off is a three-dial decision no single operator should hand-tune on every agent. Devotel Orbit's model-presets catalog packages that decision into four named bundles — Balanced, High Intelligence, Ultra Fast, and Cost Saver — each pre-wiring the transcriber, LLM, and voice, and each carrying a latency / cost / quality annotation on the Model Presets dashboard page. Picking a preset is one click; the catalog instantiates a draft voice agent pre-wired with the chosen configuration and opens it on the agent's studio page. This guide walks through why the free-form dropdown is a risk, what each preset optimises for, and how the instantiate flow composes with squads and the guardrail levers.
1. The free-form model dropdown is the failure the catalog closes
Left as an open text field, the model choice drifts. One operator picks a heavyweight reasoning model for a FAQ agent and burns budget; another picks a cheap small model for a billing-negotiation agent and burns trust. Neither picks consistently with the third operator on the same team. The failure mode is not that any single wrong choice is catastrophic — it is that production behavior drifts between agents because "which model is correct" was never a reviewed, cataloged decision. The guardrails accuracy-training comparison names the same lever explicitly: on model presets, a governed catalog beats freestyle prompting for any team larger than one operator.
2. The four named presets — what each optimises for
The catalog ships four presets on the dashboard's Model Presets page, and each card carries a curated estimated latency (P50 time to first audio byte), an estimated blended per-minute cost, and a 1–5 reasoning quality score. The metrics are curated estimates reflecting each model tier's published profile — a comparison aid at selection time, not a per-call SLA — the same curated framing the competitor preset tables use.
| Preset | LLM | Latency (p50) | Cost / min | Quality | Optimises for |
|---|---|---|---|---|---|
| Balanced | Claude Sonnet 4.6 | ~900 ms | ~$0.09 | 4/5 | The default pick — good reasoning without over-paying for calls that don't need it |
| High Intelligence | Claude Opus 4.7 | ~1400 ms | ~$0.22 | 5/5 | Deepest reasoning — multi-step negotiation, technical support, compliance or financial stakes |
| Ultra Fast | Claude Haiku 4.5 | ~450 ms | ~$0.05 | 3/5 | Lowest time-to-first-audio — high-volume triage and short, scripted flows |
| Cost Saver | Claude Haiku 4.5 | ~650 ms | ~$0.03 | 3/5 | Cheapest per-minute — very high volume, low-stakes conversations |
Two deliberate catalog boundaries matter at selection time. The bundled transcriber is always Deepgram, and the bundled voice is always Cartesia — picked because they work out of the box without a bring-your-own provider key. A preset can only bundle engines that work for every tenant on first call, so the catalog keeps the defaults dependable and leaves per-agent upgrades (an ElevenLabs voice, a BYO transcriber) to the operator afterward on the agent's own config screens.
3. How one-click instantiate works
The one-click flow works the same way as the voice-template marketplace. Pick a preset — from its card on the Model Presets page or from the preview dialog — and Devotel Orbit creates a new draft voice agent, pre-wired with the preset's model, Deepgram transcriber, Cartesia voice, sampling parameters, and a generic starter prompt. The dashboard then opens that agent's studio page, ready for your content.
The preset's job is model-and-transcriber selection, not full agent assembly. The draft still expects you to attach the knowledge base, write the real prompt, and set up phone routing — the catalog closes the recurring "which model / which transcriber / which voice" decision so the work left on a new agent is the agent's own content. Presets you can see are the ones you can instantiate, and a wrong pick is cheap to reverse: re-instantiate a different preset and discard the draft.
The draft appears once it is ready, so a preset pick is always a single deliberate action.
4. A worked example — the front door and the billing specialist
Pair presets with the agent squads orchestration post's routing structure. A tenant wires a reception-style squad where a Front door agent handles first contact and a Billing specialist handles the subscription intents (invoice, refund, payment-failed). The front door's job is greet-and-route, so it instantiates Ultra Fast — the sub-second lane answers "Hi, how can I help?" without a half-second of dead air. Billing's job is dispute resolution, so it instantiates High Intelligence — the extra reasoning depth earns more than the added latency costs, because a wrong billing answer is the failure mode that loses the customer.
The squad's routing guardrails stay intact throughout — the classifier cannot be a squad member, the fallback cannot point back at the classifier, and routing loops are rejected rather than silently cycling callers. Presets slot into that structure without touching it — the squad decides which agent handles the turn; the preset decides which bundle that agent runs.
For inbound voice the same composition applies at the phone-number layer: the number's assigned inbound agent can be the squad classifier, and the first utterance of the call is classified just enough to decide which specialist continues — with the preset already stamped on whichever agent takes over.
5. Why presets are a lever, not a cosmetic button
The guardrails accuracy-training comparison post grades the vendor field on four accuracy levers, and the "Governed model presets catalog" row is the one this guide's subject flips against the named competitors (Vapi, Retell, Bland, Synthflow, ElevenLabs, Kore.ai, Cognigy, PolyAI, Ada). The argument the row makes is the argument this guide has been running: on model presets, a governed catalog beats freestyle prompting for any team larger than one operator, and the voice-AI specialists' counter-lead on open model marketplaces is real, but it is a different cell.
The lever only exists as a lever because one-click instantiate, the curated metrics on every card, and the squad composition are all available today — the catalog is not a suggestion pinned to the dashboard, it is the decision packaged so the team stops re-deciding it on every agent. The front-door example above is what that looks like when it composes.
Frequently asked questions
What does a model preset actually change on my agent?
The preset stamps the agent's LLM (Claude Sonnet 4.6 / Opus 4.7 / Haiku 4.5 depending on the preset), the Deepgram transcriber, the Cartesia voice, and the preset's sampling parameters (temperature, max tokens). The rest of the agent — prompt, knowledge base, tools, squad routing — is the operator's own content the draft still expects.
Are the latency, cost, and quality numbers live measurements?
No — they are curated estimates per preset reflecting each model tier's published profile, shown so the operator can trade off before committing. They are a comparison aid at selection time, not a per-call SLA; per-call measurements come from the agent's own analytics once it runs.
Why are Deepgram and Cartesia the only bundled defaults?
Because they work out of the box without a bring-your-own provider key. A preset can only bundle engines that work for every tenant on first call — that is the catalog's "dependable in one click" floor. Per-agent upgrades onto other providers stay available on the agent's own configuration screens after instantiate.
Can I mix presets across a squad's classifier and specialists?
Yes, and the composition is where the catalog pays off: a cheap-and-fast preset on the classifier, a deeper preset on each specialist, with the squad's anti-loop checks still guarding the routing structure regardless of which bundle each member runs.
What if I pick the wrong preset?
Re-run instantiate with a different preset; the catalog mints a fresh draft each time and the wrong draft can be discarded. The instantiated agent is a draft by design, so the pick is cheap to reverse before the agent ever takes traffic.
The takeaway
A voice agent's model choice is a three-dial trade-off — latency, cost, quality — that no team should hand-tune per agent. The model-presets catalog packages the trade-off into four reviewed bundles with curated comparison metrics, instantiates a dependable draft in one click, and composes with squads and the guardrail levers as the pick that stays picked. The free-form dropdown was the failure; the catalog is the reviewed decision that replaces it.