Cartesia Sonic

cartesia/sonic · by Cartesia · state-space-model

Pricing — 1 offering(s)

Cartesia

About this offering: Cartesia's low-latency text-to-speech model, now sold as Sonic-3.6. It uses a state-space model (SSM) rather than a transformer and targets under 50 ms to first audio on streaming requests, which suits real-time voice agents. It supports voice cloning, WebSocket streaming and emotion and pacing control; English comes first, with growing multilingual support.

Text characters

$65.00 / 1M characters Historical 2024-04-01 → 2026-09-19
Earlier pay-as-you-go rate, before Cartesia moved to credit-based plans.
$50.00 / 1M characters Current 2026-09-20 → present

Cartesia sells credits rather than a pay-as-you-go rate, at 1 credit per character. This is the Pro plan's…

Cartesia sells credits rather than a pay-as-you-go rate, at 1 credit per character. This is the Pro plan's included rate ($5 a month for 100K credits). Overage costs $65 per 1M credits on Pro, $45 on Startup and $38 on Scale (Cartesia pricing page, checked 2026-10-11).

Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.

See how Cartesia Sonic fits into a cost-aware routing setup

Capability profile

How Cartesia Sonic rates across core capability dimensions, with the task-level evidence behind each rating.

voice naturalness Moderate

Natural for a latency-optimised model; not best-in-class on prosody richness vs ElevenLabs Multilingual v2, but strong for real-time conversational AI.

language support Moderate

English-first; multilingual support is expanding but not yet at the breadth of Azure (140+) or Google (40+).

voice variety Moderate

Library of preset voices plus voice cloning (Sonic English). Smaller voice library than ElevenLabs.

streaming latency Strong

Sub-50ms first-audio-byte via WebSocket streaming — fastest among major TTS providers. The SSM architecture enables constant-time per-token generation.

cloning support Moderate

Voice cloning from reference audio supported. Quality is solid but not at the level of ElevenLabs Professional Clone.

Benchmarks

Benchmark	Score	Config	Source
First-audio-byte latency (vendor-reported)	50 ms	Vendor-reported sub-50ms target. Not independently verified against ElevenLabs Flash v2.5 (~75ms) under identical conditions.	source ↗

Operator guidance

The first choice when streaming latency is the main constraint. Sonic costs $50 per 1M characters on Cartesia's Pro plan, more than ElevenLabs Flash v2.5's $40 per 1M pay-as-you-go (both billed per character), for a modest quality difference, but its sub-50 ms latency floor suits real-time voice agents. For batch or latency-tolerant speech, ElevenLabs or Fish Audio offer better value.

Use cases

Real-time conversational AI requiring sub-50ms voice response
Interactive voice agents where latency is the primary constraint
Low-latency TTS pipelines replacing transformers-based models
Applications sensitive to streaming latency (voice assistants, phone bots)

Limitations

English-first; limited multilingual coverage vs Azure or Google
Smaller voice library than ElevenLabs
Sold through credit-based plans, so the effective per-character rate depends on the plan and on overage
Early-stage company; pricing and feature stability less certain than established cloud providers