Compare source-linked STT, LLM, and TTS pricing verified 2026-07-22.
The calculator combines speech-to-text, language-model, text-to-speech, and infrastructure assumptions. Rates are normalized where defensible; records with host-dependent, token-billed, native-currency, or runtime pricing remain visible but are excluded from incompatible calculator totals.
| Provider | Model | Status | Input | Output | Source |
|---|---|---|---|---|---|
| OpenAI | GPT-5.6 Sol | active | $5.00 | $30.00 | Verify GPT-5.6 Sol pricing |
| OpenAI | GPT-5.6 Terra | active | $2.50 | $15.00 | Verify GPT-5.6 Terra pricing |
| OpenAI | GPT-5.6 Luna | active | $1.00 | $6.00 | Verify GPT-5.6 Luna pricing |
| OpenAI | GPT-5.4 nano | active | $0.20 | $1.25 | Verify GPT-5.4 nano pricing |
| OpenAI | GPT-5.4 mini | active | $0.75 | $4.50 | Verify GPT-5.4 mini pricing |
| Provider | Model | Mode | Status | Price | Source |
|---|---|---|---|---|---|
| OpenAI | GPT-4o Transcribe | pre-recorded | active | Token-billed | Verify GPT-4o Transcribe pricing |
| OpenAI | GPT-4o mini Transcribe | pre-recorded | active | Token-billed | Verify GPT-4o mini Transcribe pricing |
| OpenAI | GPT-4o Transcribe Diarize | pre-recorded | active | Token-billed | Verify GPT-4o Transcribe Diarize pricing |
| OpenAI | Whisper | pre-recorded | active | $0.0060/audio min | Verify Whisper pricing |
| OpenAI | GPT-Realtime-Whisper | streaming | active | $0.02/audio min | Verify GPT-Realtime-Whisper pricing |
| Provider | Model | Billing | Status | Price | Source |
|---|---|---|---|---|---|
| OpenAI | TTS-1 | characters | active | $15.00/1M chars | Verify TTS-1 pricing |
| OpenAI | TTS-1 HD | characters | active | $30.00/1M chars | Verify TTS-1 HD pricing |
| OpenAI | GPT-4o mini TTS | tokens_and_audio | deprecated | See source | Verify GPT-4o mini TTS pricing |
| Mistral | Voxtral Mini TTS | characters | active | $16.00/1M chars | Verify Voxtral Mini TTS pricing |
| ElevenLabs | Flash v2.5 | characters | active | $50.00/1M chars | Verify Flash v2.5 pricing |
Select one compatible model for each layer, enter conversation behavior and infrastructure assumptions, then review total and per-minute estimates. Fixed system/tool input and non-spoken output tokens default to zero because they are workload-specific. Public list pricing can exclude taxes, regional differences, commitments, free tiers, add-ons, cache writes, and negotiated contracts.
The calculator uses a static, source-linked catalog verified on July 22, 2026. It produces an estimate from your assumptions; it is not a quote and does not claim a fixed accuracy percentage. Taxes, negotiated rates, free tiers, add-ons, caching, regional differences, and request minimums may change the invoice.
The catalog covers the providers shown on the LLM, STT, and TTS comparison pages. Entries that cannot be normalized honestly—such as token-billed transcription, runtime-priced community models, or native-currency rates—remain visible in the catalog but are excluded from deterministic calculator totals.
Yes. Keep conversation assumptions fixed, switch one component at a time, and compare the breakdown. Only rank rows with compatible modes, units, regions, and tiers; a batch STT price is not a realtime substitute.
Key inputs include conversation length, speech share, words and turns per minute, accumulated LLM context, model prices, and optional hosting cost. Taxes, free tiers, cache hits, regional pricing, minimum billing increments, and add-ons must be checked separately.
Yes. The voice AI cost calculator is free to use without registration. You can run calculations, export results, and create share links without a site account.
Every catalog row shows its verification date and official source. The current release was checked on July 22, 2026. Always open the source link before making a purchase because providers can change pricing between site releases.
Yes. CSV exports include provider IDs, official source URLs, catalog date, assumptions, and unrounded results. Share links preserve the calculation state in a read-only view.
The latency panel is an editable budget, not a live provider benchmark. Enter measurements from your own client, regions, network path, STT, LLM, TTS, and audio buffers to estimate end-to-end response time.
Start with context management, measure actual input and output tokens, test smaller current models, use caching only when your prompt pattern earns cache hits, and compare the correct STT mode and TTS plan. The calculator deliberately keeps input and output token prices separate.
Measure endpointing, network transit, LLM time to first token, sentence aggregation, TTS time to first audio, and client audio buffers separately. Co-locate services where possible, stream partial results, and validate with production traces rather than generic benchmark numbers.
The trade-off is workload-specific. Use the latency panel to enter measurements from your own regions and providers, then compare the resulting estimate with actual invoices. Avoid universal cost or latency targets that are not tied to a measured workload.