Cartesia Sonic-3.6: Real-Time Voice Agent Guide

The short answer

Cartesia Sonic-3.6 is a beta text-to-speech model for developers building real-time voice agents and interactive audio applications. Cartesia’s launch material positions it as a sub-90 ms time-to-first-audio model, while its current pricing page lists Sonic-3.6 across Free, Pro, Startup, Scale, and Enterprise plans. Artificial Analysis currently ranks Sonic 3.6 at the top of its provider-voice text-to-speech comparison with a 1,285 Elo score and lists a representative price of $49 per million characters.

The practical takeaway is not that Sonic-3.6 automatically replaces every TTS provider. It is that teams building synchronous agents should evaluate it as a low-latency hosted API, especially when natural turn-taking, expressive controls, voice cloning, and predictable streaming matter more than self-hosted weights. The model is commercial and hosted; it is not an open-weight model you can download and run on your own GPUs.

What launched

Cartesia introduced Sonic-3.6 alongside Ink-2, its streaming speech-to-text model, as part of a voice stack aimed at interactive intelligence. The company says the two models are designed together for the real-time loop: speech recognition, turn detection, language-model response, and speech generation.

Sonic-3.6 is the text-to-speech side of that stack. The official launch page highlights sub-90 ms TTS latency, expressive output, and a state-space-model architecture. Cartesia also says Sonic supports more than 40 languages, while its launch messaging specifically describes 44-language coverage. Because model availability and language support can change during beta, production teams should verify the current model matrix before committing to a locale.

The release is important for agent builders because a voice agent is judged by the complete conversation, not by a standalone audio sample. Time-to-first-audio affects whether the agent feels responsive after a user stops speaking. Prosody, pronunciation, and non-verbal cues affect whether the response sounds usable. The API’s streaming behavior and concurrency limits affect whether the system survives real traffic.

Visual — streaming voice-agent path. User audio → streaming STT → turn detection → agent/LLM → policy checks → Sonic-3.6 → buffered audio. This original flow is an editorial architecture visual, not an official Cartesia screenshot.

The production architecture

A useful baseline architecture keeps the TTS provider behind an adapter rather than spreading provider-specific request formats through the application.

The model should not be responsible for authentication, authorization, prompt policy, tool permissions, or sensitive-data filtering. Those belong in the agent runtime. The TTS adapter should receive an already-approved transcript and return audio chunks, status, provider request IDs, and timing measurements.

This separation also makes fallback easier. For example, an application can use Sonic-3.6 for the primary path, route a failed request to another provider, and preserve the same agent-level interface. The fallback decision should account for voice identity and language; switching providers can be technically successful but perceptually jarring.

For a broader systems view, compare this design with the production guide to OpenAI realtime voice agents and the OpenTelemetry guide for generative-AI observability.

API shape and an implementation starting point

Cartesia’s current documentation exposes a bytes endpoint at POST https://api.cartesia.ai/tts/bytes. The documented request includes a model ID, transcript, voice ID, output format, locale, normalization options, pronunciation dictionary, and generation settings such as speed and volume.

The following request mirrors the documented endpoint and header shape. Sonic-3.6 is in beta, so verify the exact accepted model identifier and account access in the current Cartesia model documentation before running it in production.

1curl --request POST \\
2  --url https://api.cartesia.ai/tts/bytes \\
3  --header "Authorization: $CARTESIA_API_KEY" \\
4  --header "Cartesia-Version: 2026-08-14" \\
5  --header "Content-Type: application/json" \\
6  --output reply.wav \\
7  --data @- <<'JSON'
8{
9  "model_id": "sonic-3.6",
10  "transcript": "Your appointment is confirmed for tomorrow at ten AM.",
11  "voice": "db6b0ed5-d5d3-463d-ae85-518a07d3c2b4",
12  "output_format": {
13    "container": "wav",
14    "encoding": "pcm_s16le",
15    "sample_rate": 44100
16  },
17  "locale": "en",
18  "generation_config": {
19    "volume": 1,
20    "speed": 1
21  }
22}
23JSON

Do not hard-code the example voice in a real application. Store voice IDs in configuration, validate that the selected voice is approved for the requested use, and keep the API key in a secret manager. The public documentation example currently uses sonic-3.5; that is a reminder to check beta documentation rather than assuming every example has already been updated for Sonic-3.6.

For streaming agents, measure the time from the final user-audio frame to the first playable audio frame. A provider’s model-level time-to-first-audio claim is not the same as end-to-end latency. WebSocket setup, network distance, speech endpointing, language-model generation, audio decoding, and client buffering all contribute to the user experience.

Latency: what the claim does and does not mean

Cartesia states sub-90 ms TTS latency on its launch materials. MarkTechPost reports the same figure as a vendor-stated time-to-first-audio claim and explicitly distinguishes it from a measured end-to-end round trip. That distinction is essential.

A production benchmark should record at least:

• final-input-to-first-audio latency;
• first-audio-to-complete-response latency;
• audio underruns and reconnects;
• p50, p95, and p99 latency by region;
• latency by transcript length and language;
• the percentage of responses that require a fallback;
• cost per completed conversation minute.

Run the test from the same regions and network conditions your users will use. A local benchmark from a developer laptop can make a hosted service look faster than it is for a globally distributed customer-support product.

Pricing and capacity

Cartesia’s current pricing page lists Sonic-3.6 usage through credits and subscription plans rather than a simple public per-character rate. The page shows these representative tiers:

Artificial Analysis presents a normalized comparison price of $49 per million characters for Sonic 3.6. That number is useful for cross-provider comparison, but it should not be confused with Cartesia’s subscription invoice. The two figures use different accounting models, so finance teams should calculate expected monthly characters, audio minutes, concurrency, cloning needs, and overages against Cartesia’s own plan terms.

The pricing page also lists instant voice cloning on Pro and professional voice cloning on higher tiers, with enterprise features such as DPAs, BAAs, SSO, and custom concurrency limits. Those features may matter more than the headline model price for regulated deployments.

The right cost comparison is usually cost per successful task, not cost per million characters. If a faster voice reduces interruptions, retries, abandoned calls, or human transfers, a higher unit price can still lower the cost of a completed workflow.

How Sonic-3.6 compares with the alternatives

Artificial Analysis currently lists Sonic 3.6 at 1,285 provider-voice Elo, ahead of Qwen-Audio-3.0-TTS-Plus at 1,240, Simba 3.2 at 1,240, Luna TTS at 1,221, and Gemini 3.1 Flash TTS at 1,212. The ranking is a snapshot of an evolving evaluation, not a universal verdict for every language, voice, or application.

For model-selection context, the Grok 4.6 frontier-agents guide covers a different part of the agent stack: reasoning and tool use rather than speech synthesis. For retrieval-heavy voice workflows, the agentic RAG guide is a useful companion.

Security and privacy checklist

Treat voice as sensitive data. Before enabling production traffic:

1. Confirm retention, deletion, regional processing, and training-use terms.
2. Send only the approved transcript to TTS; do not pass hidden system prompts or tool credentials.
3. Add an allowlist for voice IDs and languages.
4. Rate-limit requests and cap transcript length to control spend.
5. Log provider request IDs and timing, but redact the transcript unless retention is justified.
6. Require explicit consent for voice cloning and document the owner of the source recording.
7. Use a fallback that preserves safe-content filtering and does not silently change a regulated disclosure.
8. Test pronunciation of account numbers, dates, medication names, URLs, and confirmation codes.

The API key should never be exposed to a browser client. Put the provider call behind a server-side service or an edge function with scoped secrets. If the product needs client-side low-latency audio, issue short-lived session credentials only if the provider supports that model safely; otherwise proxy the request.

Common integration failures

The request returns an unknown model error. Sonic-3.6 is beta and the current public docs may still show Sonic-3.5 in examples. Confirm account access, the model catalog, API version, and exact model ID.

The voice sounds slow even though the model is fast. Check speech endpointing, language-model delay, output buffering, and the client’s playback queue. TTS speed is only one part of conversational latency.

Pronunciation is inconsistent. Use the pronunciation dictionary and normalization controls rather than adding ad hoc spelling hacks to every prompt. Build regression audio tests for names and domain vocabulary.

Costs exceed the estimate. Track characters, audio minutes, concurrency, retries, and voice-agent charges separately. A retry storm can dominate a normal unit-cost calculation.

The beta output changes. Pin API versions where supported, keep golden audio samples for acceptance testing, and make provider/model selection configurable.

FAQ

Is Sonic-3.6 open source?

No. Current evidence describes it as a hosted commercial model available through Cartesia’s API and plans. Do not plan an offline or self-hosted deployment around it.

Is sub-90 ms the end-to-end response time?

No. It is a vendor-stated model-level latency claim. Measure from the end of user speech to the first audio byte your client can play.

Is Sonic-3.6 available for free?

Cartesia’s pricing page lists a Free plan with a limited monthly Sonic-3.6 allowance. Production commercial use and additional capabilities depend on the selected plan.

Should every voice agent use Sonic-3.6?

No. Choose it when low latency, expressive speech, and hosted operations fit the product. Compare alternatives when you need open weights, a specific language, self-hosting, a different retention policy, or lower cost at your workload.

What should I test first?

Start with a 24-hour replay set of real, consented transcripts. Measure latency percentiles, pronunciation accuracy, interruption behavior, audio failures, user preference, and cost per successful task before changing the default provider.

Bottom line

Sonic-3.6 is a credible candidate for real-time voice agents because its launch is paired with a production API, documented controls, plan-level capacity information, and an independent evaluation that currently places it at the top of Artificial Analysis’ provider-voice leaderboard. The strongest case is interactive systems where response timing and natural delivery directly affect task completion.

The responsible implementation is verification-first: use a server-side adapter, measure end-to-end latency, keep a fallback, test every important locale and pronunciation, and calculate cost from real conversations. Treat the beta model identifier, availability, and leaderboard position as changeable facts rather than permanent guarantees.

Sources and visual credits

• Cartesia: Introducing Sonic-3.6 and Ink-2 — official launch and architecture/latency claims.
• Cartesia pricing — official plans, credits, concurrency, and Sonic-3.6 allowances.
• Cartesia API documentation: Text-to-Speech (Bytes) — official endpoint and request shape.
• Artificial Analysis: Sonic 3.6 — independent quality, price, and speed comparison.
• MarkTechPost: Cartesia Ships Sonic-3.6 — independent technical coverage and beta/deployment context.
• The Mermaid request-flow diagram is an original editorial visual by Essam A. Mamdani. No official product screenshot is implied.