Changelog

API updates, new endpoints, and breaking changes.

v1.3.1vyasNovember 2026
  • +Vyas now runs on the same servers as voice agents. Its context window is 16,384 tokens, prompt and answer together, and GET /v1/models reports it. A longer prompt is refused with context_length_exceeded, and max_tokens is reduced to what fits.
v1.3.0vyasOctober 2026
  • +Vyas, the Sonex Labs language model, is now on the API, OpenAI compatible. Point any OpenAI client library at https://api.sonexlabs.com/v1 with your Sonex API key and use the model vyas.
  • +POST /v1/chat/completions takes OpenAI's chat completions request, including tools, response_format and streaming. GET /v1/models lists the model and its context window.
  • +Billed per input and output token, as reported in usage. Requests refused before the model runs are not billed.
  • +Context window of 65,536 tokens, prompt and answer together, and the same rate limits as every endpoint: 5 requests per second and 120 per minute. Errors use OpenAI's error shape, so OpenAI client libraries raise the matching exception.
v1.2.0speech-to-textSeptember 2026
  • +Speech-to-Text API: transcribe audio with Valmiki, our transcription engine. 100+ languages with automatic detection and punctuation.
  • +POST /v1/transcribe takes a finished recording, as a raw body or multipart upload, and returns the transcript with its duration. WAV, MP3, FLAC, OGG, Opus, WebM, AAC and M4A, up to 10 MB and 10 minutes.
  • +WS /v1/transcribe/stream takes 16 kHz mono PCM audio and returns partial and final transcripts as they are recognised. Browsers that cannot set a WebSocket header may pass the key as api_key in the query string.
  • +Transcription is billed per second of audio, because a transcriber charges for the time it is listening whether or not anyone is speaking. A recording is billed for its own duration; a stream for as long as it is open. Speech-to-text credits are drawn before your wallet.
  • +Audio we reject (unsupported format, over 10 MB, over 10 minutes, unreadable) is never billed. An Idempotency-Key makes a retry safe: the transcript comes back again with seconds_billed of 0.
v1.1.0callsSeptember 2026
  • +Call history API: GET /v1/calls returns your calls newest first, with filters for agent, status, direction, phone number and date range. Paginated with a cursor rather than page numbers, so a page never repeats or skips a call while new ones arrive.
  • +Call detail: GET /v1/calls/{call_id} returns the models used, the itemised cost, the transcript and every tool the agent invoked with its arguments, result and duration. Use the include parameter to choose which sections are returned.
  • +Transcripts: GET /v1/calls/{call_id}/transcript returns the transcript on its own, as turns and as a single block of text.
  • +Recordings: GET /v1/calls/{call_id}/recording redirects to the audio, or returns a time limited link as JSON with redirect=false. Links are valid for 15 minutes.
  • +Voice previews now have a stable address. preview_url on each voice points at GET /v1/voices/{voice_id}/preview, which needs no API key and can be used directly as the src of an audio element.
  • +Every call endpoint is scoped to the tenant that owns the API key.
v1.0.0launchJune 2026
  • +Public API launched: Text-to-Speech, Voices. Agents, Agent Tools, and Call Recordings coming soon.
  • +Pāṇini TTS: 250+ languages, streaming synthesis over HTTP.
  • +Voice cloning from as little as 10 seconds of reference audio.
  • +Agent builder with tool library (coming soon): attach webhook or function tools to any agent.
  • +Bearer token authentication with per-key scoping.