Estimated per-request cost for a completed TTS request, in US dollars
(Request.cost, a Decimal, as for chat/embeddings).
No TTS provider self-reports cost the way OpenRouter's chat completions do (there's
no usage object at all in an /audio/speech or xAI /tts response) — so unlike
Completion.extract_usage/1, this can't extract a real figure from the response. It
has to come from a hand-maintained rate table instead.
Rates as of 2026-07-16 — verify against each provider's current pricing page before trusting for billing-critical use. These will drift; this table needs periodic manual updates, same as any other hardcoded rate table. Sources:
- Mistral (
voxtral-mini-tts-2603): $16 / 1M characters — https://mistral.ai/pricing/api/ - xAI (Grok TTS — one flat rate regardless of model, since xAI TTS has no per-model pricing tiers): $15.00 / 1M characters — https://docs.x.ai/developers/pricing (corrected 2026-07-16: originally shipped as $4.20/1M in 0.15.0, sourced from an OpenRouter resale listing rather than xAI's own direct-API rate — this module only ever calls xAI's own API, so that number was never the applicable one)
- OpenAI (
gpt-4o-mini-tts): $0.60 / 1M input text tokens + $12 / 1M audio output tokens — https://developers.openai.com/api/docs/models/gpt-4o-mini-tts
Mistral and xAI bill per character, and input_chars is exact (String.length/1
on the request text), so their cost is exact given the rate is current. OpenAI bills
per token on both legs, but /audio/speech returns no token counts to read back —
estimating a token count from characters (further split into an input-text leg and
an audio-output leg) would compound two guesses, so this uses the commonly-cited
blended rate (~$0.015/minute of generated audio) applied to a duration estimated
from audio_bytes at the assumed bitrate instead — one estimate, not two. Treat it
as approximate wherever it's displayed.
Summary
Functions
Estimated cost in dollars for a completed TTS request, rounded to the cost
column's 18 places, or nil if provider isn't in the rate table — same as
no cost being available today, not a new failure mode. input_chars and
audio_bytes are exactly what PhoenixKitAI.speak/3 already records in the
request's metadata.
Functions
@spec cost(String.t(), non_neg_integer(), non_neg_integer()) :: Decimal.t() | nil
Estimated cost in dollars for a completed TTS request, rounded to the cost
column's 18 places, or nil if provider isn't in the rate table — same as
no cost being available today, not a new failure mode. input_chars and
audio_bytes are exactly what PhoenixKitAI.speak/3 already records in the
request's metadata.