* feat(mistral): add Voxtral audio models
Mistral ships a full audio line that the catalog does not cover yet:
transcription, text-to-speech and an instruct model with native audio input.
- voxtral-mini-latest: audio to text transcription
- voxtral-mini-tts-latest: text to audio, zero-shot voice cloning, 9 languages
- voxtral-small-latest: audio+text to text, tool calling, 32k context
The two first ones intentionally omit the [cost] block: transcription bills per
MINUTE of audio (\$0.003/min) and synthesis per CHARACTER (\$16 per 1M chars),
neither of which the token-based schema models. Same treatment as the existing
Whisper entries, e.g. providers/groq/models/whisper-large-v3-turbo.toml.
Voxtral Small does carry token pricing for its text side; its audio input bills
per minute (\$0.004) and is documented in the file header.
Sources are cited as a leading comment block in each file, per AGENTS.md.
Validated with bun validate.
* fix(mistral): align Voxtral Mini entries with the live API ids
voxtral-mini-latest resolves to voxtral-mini-2602, not the 25-07
Transcribe card the entry was named and dated after. Date the entry on
the revision it points at, matching mistral-small-latest, and drop the
product word absent from the API id. Note the Bedrock Voxtral Mini 3B
entry as a distinct product surface to prevent the same confusion.
Name the TTS entry after its own id for consistency.
Nourrisse Florian
44fecad3ac
