How this index works

What counts as a realtime API

Anything that streams audio or video while a conversation is happening: voice-to-voice models, streaming speech-to-text and text-to-speech, the platforms and phone carriers that move the audio, talking avatars, realtime video models, and open models you can run yourself. Batch-only APIs are left out.

Where the data comes from

Each entry was researched against the vendor's own pricing page, docs and changelog, plus price-list APIs where the public page hides numbers (Azure and AWS). Every page links the sources that were used. When something could only be confirmed by a third party, or not at all, the page says so in a "Not fully verified" note and the confidence drops.

The whole set was checked on Oct 10, 2026. Each API shows its own check date, and anything older than 60 days is flagged as possibly out of date.

Per-minute estimates

Vendors bill by audio tokens, characters, UTF-8 bytes, open connection time or flat minutes, so a single "price per minute" always involves assumptions. Every estimate has its working shown on the provider page. For token-billed voice-to-voice models, the low figure is a single short turn and the high figure is a 10 minute, 50 turn call where the whole history is billed again on each turn with no cache hits.

Confidence

46 entries are high confidence, 79 entries are medium confidence, 15 entries are low confidence. High means every price line was read on an official page. Low means key numbers came from third parties or could not be found.

Code samples

Quick starts were written from current docs and kept short on purpose. SDKs in this space change monthly, so check the version you install. Never ship a long-lived API key to a browser; every major provider has a short-lived token for that.

Corrections

Every API page has a "Report outdated info" form. Reports go to a review queue and corrections are applied by hand.

Data

The full data set is available as JSON for your own comparisons.

Who made this

Built by Anri while building voice features for other projects, after one too many surprise invoices. No vendor paid to be listed and there are no affiliate links.