Moonlit terraces shaped like speech waveforms under a night sky
The Economics

Priced for every night sky.

No plans, no slabs, no monthly minimums. Billed per second of audio: feather for high-volume work, starling for studio quality and cloning.

RATE CARD · 2026
SERVING · FEATHER + STARLING
01 · The Rate

One number to remember.

0.19¢ / 1,000 characters · feather

All 23 languages and every voice included, at the same rate. At published list prices that is 9× to 53× cheaper than the leading commercial voice APIs: press play below and watch the bill. Three dollars buys roughly 1,750 minutes of speech. Billed by the second, no minimums, cancel any time.

feather · per minute of audio$0.0017 · 15 paise
feather · per hour of audio$0.10
feather · per 1M characters$1.90 · ~1,100 min
starling · per 1M characters$9.00 · studio + cloning
Free, renewed monthly10,000 chars
02 · The Field

Where the rate sits among providers.

Published list prices, normalised to the same two units. The difference is compute: everyone else rents GPUs, the Observatory runs on ordinary CPUs.

Provider$/1K chars$/min audioCompute
CelestLabs feather$0.0019$0.0017CPU
OpenAI tts-1$0.015$0.015GPU
Amazon Polly$0.016$0.016GPU
Cartesia Sonic$0.036$0.036GPU
ElevenLabs$0.150$0.100GPU
03 · The Same Line

Seven engines. One sentence. Watch the bill.

Each row is a real recording of the same line from that provider's model. The meters count up in lockstep with playback: each number is that model's price per million characters, not the cost of this clip.

FIG. 01 · seven engines, published list prices per 1M characters, each paired with the model you hear.

CelestLabsfeather · CPU
$0.00/ 1M chars1× · you
Smallestlightning-v3.1
$0.00/ 1M chars
Deepgramaura-2
$0.00/ 1M chars16×
OpenAItts-1-hd
$0.00/ 1M chars16×
Sarvambulbul:v3
$0.00/ 1M chars17×
Cartesiasonic-3.5
$0.00/ 1M chars21×
ElevenLabseleven v3
$0.00/ 1M chars53×

Published list prices per 1M characters · verified June 2026

04 · Why the Price Holds

Engineering, not arithmetic.

This is not a loss leader, an introductory rate, or a thin reseller margin. It is what synthesis costs when the whole path runs on commodity CPUs.

OBS. A · NO GPUS
No GPUs anywhere in the path.
A 117M-parameter distilled model on an ONNX runtime, streaming at 6.1× realtime (RTF 0.18) on an ordinary 8-core CPU. No CUDA, no accelerator surcharge passed along to you.
OBS. B · EVERY REGION
The rate holds in every region.
CPUs exist in every cloud region on earth, so the price does not change with geography. Serve your users where their data must live, at the same 0.22¢ per 1,000 characters.
OBS. C · ON-PREM
Small enough to bring home.
The serving stack fits in roughly 500 MB and a single process, and cold-starts in about two seconds. It runs on-premise and air-gapped, wherever the Observatory cannot reach.

The full methodology, model card, and per-machine latency table live on the research page.

Hear it first

Numbers are nice. Voices are better.