Sonic 3.6 key stats
- Latency (measured)
- 101 ms1
- Vapi streaming benchmark (50 trials per model) (checked 2026-10-26) Median of 50 sequential WebSocket trials after two discarded warm-ups, using the Clara professional clone and raw 24 kHz PCM. Measured on a local macOS machine; excludes connection setup and includes network RTT.
- docs.cartesia.ai/build-with-cartesia/tts-models/latest (checked 2026-10-26)
- cartesia.ai/pricing (checked 2026-10-26) The displayed $50 per 1M characters is the effective rate of Pro included credits ($5/month for 100K), assuming full use, rather than an overage rate. Startup included credits are about $39.20/1M ($49 for 1.25M); Scale about $37.38/1M ($299 for 8M). Overages cost $65/$45/$38 per 1M credits on Pro/Startup/Scale. TTS uses approximately 1 credit per character; preprocessing can affect usage (docs.cartesia.ai/pricing). Professional voice clones require Startup or higher; custom Enterprise rates may differ.
- docs.cartesia.ai/build-with-cartesia/tts-models/latest (checked 2026-10-26) Snapshot release.
Background
Sonic 3.6 is Cartesia's current flagship model. The stable snapshot released on August 27, 2026 supports 44 languages and works with professional voice clones created on older models. This entry pins that dated snapshot for reproducibility.
Sources: docs.cartesia.ai
At a glance
English arena clips use the same four licensed source voice clones and 20 frozen scripts as the rest of the field. In our separate 50 trial WebSocket benchmark, this snapshot returned first audio in a median of 101 ms, including network time from the benchmark machine.
Sources: docs.cartesia.ai
Frequently asked questions
- How is Sonic 3.6 tested on the Humanness Index™?
- Listeners hear Sonic 3.6 against another model in a blind head to head round, both voices reading the same customer support prompt from the same cloned source voice, and they pick whichever sounds more human. Its Humanness score derives purely from those votes.
- How fast is Sonic 3.6?
- Our October 26, 2026 benchmark measured a median of 101 ms from request send to first audio over an established WebSocket connection, using the Clara professional clone and raw PCM. All 50 measured trials succeeded after two discarded warm-ups. This includes network RTT and does not measure a full voice agent turn.
Find the most human-sounding voice for your agent.
Compare the models in blind tests, read the methodology, or get in touch.
Build a TTS model? Add yours to the Index.
