Build phone and web voice agents on any speech-to-text, language, text-to-speech or speech-to-speech model. The platform is a flat $0.010 per minute. Models cost the provider's list price plus a 15% fee you can see on every receipt, or $0 with your own keys.
500 free minutes every month, on every model, no card needed.
Models from Deepgram, AssemblyAI, Soniox, Speechmatics, OpenAI, Anthropic, Google, Cerebras, Cartesia, Inworld, ElevenLabs, Rime, Hume AI, Speechify, xAI, Mistral and more, each with its price and a link to the page it came from.
| Line | USD |
|---|---|
| STT · Deepgram Flux (English) verified | $0.0065 |
| LLM · GPT-5 mini verified | $0.0024 |
| TTS · Inworld TTS-2 Flash verified | $0.0090 |
| Managed keys +15% ($0 with yours) | $0.0027 |
| Voiceflint platform | $0.010 |
| TOTAL, WEB CALL | $0.031 |
Assumes 6 turns/min, 3k prompt tokens/turn with 70% cache hits, agent speaking half the time.
Most platforms charge a per-minute fee that is larger than the models underneath it, then sell concurrency and compliance on top. Voiceflint charges one fee and shows the rest.
Open the calculatorPer minute, US phone call, Balanced stack (Deepgram Flux, GPT-5 mini, Inworld TTS-2 Flash). Vendor prices read on 2026-09-08; concurrency and compliance add-ons excluded. Full table on the pricing page.
Callers start talking over an agent at about 800 ms. A well-chosen cascade answers in 600 to 900 ms, and every turn on Voiceflint is recorded as these four numbers, so you can see which model to change when it gets slow. Each agent runs in US East, US West or EU West, next to its callers and its model providers, at the same price.
Most platforms still default to a 2025 voice that now sits at #42 on the arena. These are the stacks we recommend today, and why. Change any part of them.
Flux fuses STT and end-of-turn (one fewer hop); GPT-5 mini is the cheapest model with reliable tool calling; Inworld TTS-2 Flash is top-10 arena quality at 20 ms first byte. About $0.02/min in models, 4-6x below Retell or Bland.
Soniox is the cheapest and broadest streaming STT with strong accuracy (249 ms, 1.29% semantic WER); Gemini Flash leads on non-English; Inworld TTS-2 covers 200+ languages at one price; Smart Turn handles 23 languages.
Flux eager end-of-turn starts the LLM before the turn is confirmed; gpt-oss-120b on Cerebras streams ~3,000 tok/s; Rime Mist v3 has 37 ms P50 time-to-first-audio. Faster than any hosted speech-to-speech model measured by Artificial Analysis.
AssemblyAI streaming is $0.15/hour with fused end-of-turn; GPT-5 nano handles FAQ and routing; Soniox TTS v2 is #13 on the arena at $14/1M chars. About $0.012/min in models, roughly 5x cheaper than Retell's infrastructure fee alone. Good for IVR replacement, not complex tool flows.
Cartesia Sonic 3.6 leads the arena by ~30 ELO; Claude Sonnet 5 is the best price/quality Claude with the strongest tool use; Universal-3.5 Pro is the only streaming STT with contextual awareness and real-time diarization.
Gemini Live is the cheapest hosted speech-to-speech ($0.005 in / $0.018 out per minute) with 0.63 s time-to-first-audio. Note: Daily's Feb 2026 benchmark found cascades still beat S2S on multi-turn task completion.
Each call shows the latency waterfall per turn and the exact cost per line item at the provider's published rate. Test calls get the same treatment as production. If a number is an estimate, it says so.
| STT · Deepgram Flux · 4.2 min | $0.0273 |
| LLM · GPT-5 mini · 74k tokens, 68% cached | $0.0178 |
| TTS · Inworld TTS-2 Flash · 2,460 chars | $0.0369 |
| Managed keys (+15%) | $0.0123 |
| Telephony · Voiceflint number | $0.0504 |
| Voiceflint platform | $0.0420 |
| TOTAL · $0.040 / MIN | $0.1867 |
Start from a benchmarked preset or choose each model. The receipt and the latency estimate update as you go.
Talk to your agent over WebRTC in one click. Every test call gets the same receipt and waterfall as production.
Attach a Voiceflint number at $0.012 per minute or bring your own SIP trunk from Twilio, Telnyx or anyone else.
Sign up, pick a stack, and talk to your agent in the browser within a minute. Pay only when you pass the free minutes.