Key facts
| Language layer | Chat completions, streaming and JSON mode are live on Plugsky |
| Audio endpoints | Speech-to-text and text-to-speech are coming soon; bring your own STT/TTS today |
| Router model | plugsky-fusion escalates per turn across tiers (live) |
| Intent routing | JSON mode for intent and slot output on cheap tiers |
| Pricing | Flat monthly self-serve plans with no per-token charges on self-serve |
| Free tier | plugsky-micro and plugsky-lite on the free plan, no card required |
| History control | Keep rolling summaries rather than full transcripts |
| Models | 30+ models; routine spoken turns on cheap tiers, hard requests escalated |
TL;DR
- Speech costs are separate: budget STT and TTS alongside the language layer.
- Short spoken answers are cheaper and better; cap them by design.
- Extract intent with JSON mode on a cheap tier before generating replies.
- Summarize history instead of replaying transcripts.
- Cache frequent intents and responses where personalization allows.
How it works, step by step
- Measure the three cost layers separately: recognition, language processing, synthesis.
- Cap spoken reply length in the prompt and in max_tokens.
- Route transcript normalization and intent extraction to cheap tiers with JSON mode.
- Generate replies on a tier matched to the intent's complexity.
- Keep a rolling summary rather than resending the full transcript each turn.
- Cache common intent-response pairs where personal data is not involved.
- Track cost per completed call alongside recognition accuracy and task completion.
Try it yourself
Open the LLM API cost calculator →
Where voice spend goes
Voice assistants pay three vendors or layers: speech recognition, the language model, and speech synthesis. Plugsky covers the language layer today, with audio endpoints coming soon, so recognition and synthesis remain your own line items to measure and optimise. Teams often tune the model layer while recognition retries and verbose synthesis quietly dominate.
Within the language layer, cost follows turn length and history. Spoken answers should be short by design — one or two sentences — and conversation history should be summarized rather than replayed, because a spoken session accumulates transcript quickly.
Text tiering for spoken turns
Speech adds a normalization step before anything else: punctuation, casing, numbers and likely misrecognitions. That work is cheap and should stay on cheap tiers, as should intent detection and slot extraction with JSON mode.
- Cheap tier: transcript cleanup, intent and slot extraction, short factual replies.
- Mid tier: troubleshooting steps and multi-turn task guidance.
- Strong tier: complex requests, complaints and anything requiring judgement.
- Cached: common intents and responses where personalisation is not required.
Latency, retries and measurement
Voice has a hard latency constraint, and retries are visible to the caller as dead air. A timeout that fails into a holding phrase and a human handoff is better than a long retry loop that costs more and sounds worse. Design the failure path deliberately.
Measure cost per completed call, recognition accuracy, intent accuracy and task completion. A cheap language layer paired with high recognition retries is a false saving; fix the layer that is actually failing. Start on the free plan with plugsky-micro and plugsky-lite for the language layer, then evaluate stronger models for complex calls during the 14-day full-access trial; plans are on the live pricing page.
Honest comparison
| Voice cost lever | Routed voice assistant | Strong model every turn | Unbounded spoken turns |
|---|---|---|---|
| Intent extraction | Cheap tier with JSON mode | Frontier price per turn | Same |
| Routine replies | Cheap tier, capped length | Frontier price per reply | Verbose |
| Complex requests | Escalated to strong tier | Native strength | Weak replies |
| History | Rolling summaries | Full transcript replay | Full transcript replay |
| Evidence | Cost per completed call | Token totals | Token totals |
Frequently asked questions
Does Plugsky handle speech-to-text and text-to-speech?
Not yet — audio endpoints are coming soon. Plugsky provides the language layer with live chat completions, streaming and JSON mode, and you connect STT and TTS providers of your choice.
Why do voice assistants cost more than chatbots?
They add recognition and synthesis costs around the language layer, and spoken sessions accumulate transcript quickly. Long replies and replayed history compound within the model layer.
How do I keep replies short?
Prompt for one or two spoken sentences, cap max_tokens, and avoid lists. Short answers are also better for the caller, so cost and quality align.
Should intent detection use a strong model?
No. Intent and slot extraction is structured, short work that cheap tiers handle well with JSON mode. Reserve strong models for complex requests and escalations.
How do I manage conversation history?
Keep a rolling summary of confirmed facts and pending actions rather than resending the transcript. Voice sessions generate text fast, so history discipline matters more than in chat.
Can I cache responses?
Yes, for common intents where personalisation is not required and freshness allows. Cache at the intent level rather than the raw transcript.
How do I evaluate cost?
Track cost per completed call alongside recognition accuracy, intent accuracy and task completion. Optimising the language layer while recognition retries dominate is a false saving.
Can I start for free?
Yes. plugsky-micro and plugsky-lite are on the free plan with no card, and the 14-day full-access trial covers stronger models for complex calls.