Changelog.

Every notable change to the API, model, and SDKs. Breaking changes are versioned (currently feather) and earlier major versions stay available for at least twelve months.

2026-06-25 · Starling, cloning, voice design & emotion

• New: premium model starling, instant voice cloning, voice design, 13 expression tags, and 48+ languages (all major Indic + global), still CPU-served. See Starling.
• New: custom-voice API, POST/GET/DELETE /v1/voices to enroll, list, and remove cloned/designed voices; reuse them by voice_id.
• New: /v1/synthesize and /v1/stream accept voice_id, ref_audio, ref_text, instruct, and language for Starling.
• Pricing: Starling is $0.012 / 1K characters ($9 / 1M), one price for one-shot and realtime.

2026-06-14 · live streaming demo, docs site launch

• New: sentence-by-sentence streaming is live in the demo, audio starts playing before the full text is rendered.
• Roadmap: the streaming API for your own key, HTTP stream: true on POST /v1/audio/speech and the full-duplex WebSocket at wss://api.celestlabs.ai/v1/stream, is documented and rolling out; not yet on the current infrastructure. See streaming docs.
• New: developer docs site at /docs with REST, streaming, SDK, and error reference.
• New: ephemeral-token endpoint POST /v1/ephemeral_tokens for safe browser-side use.
• Changed: single global USD price ($0.003 / min) replaces market-specific pricing.

2026-05-22 · Distilled model refresh

• New: 117M-parameter student model swaps in seamlessly under model: feather. ~22% lower compute, same audible quality. No client changes needed.
• Improved: Hindi and Spanish prosody on long-form narration after two additional fine-tuning passes.

2026-03-01 · First public beta

• New: REST endpoint at /v1/audio/speech with seven voices and twelve languages.
• New: OpenAI-compatible client surface.