# Cartesia — The Stanford SSM-research spinout behind Sonic (TTS) and Ink (STT) — Line, its open-source voice-agent SDK, connects to 100+ LLMs and starts at a genuine $0/month, but the agent layer itself isn't self-hostable.

> Source: The Agents Index — https://theagentsindex.com/cartesia (structured, researched, re-verified)
> Facts last verified: 2026-08-23

Cartesia is an AI voice platform built around State Space Models (SSMs), a research lineage its founding team helped pioneer at the Stanford AI Lab before founding the company in September 2023. It ships three products. Sonic-3.6 is the current real-time text-to-speech model, with laughing and emotional control across 40+ languages. Ink-2 is the streaming speech-to-text model. Line is the voice-agent development platform that sits on top of both. The company has raised $191 million across three rounds. A $27M seed (December 2024, led by Index Ventures), a $64M Series A (March 2025, led by Kleiner Perkins), and a $100M Series B in October 2025, led by Kleiner Perkins with Index Ventures and Lightspeed returning and NVIDIA joining as a new strategic investor. The Series B was announced alongside Sonic-3, a 42-language TTS model Cartesia said was preferred over competitors by 62% of blind human testers, and a successor Sonic-3.6 now sits on the pricing page. Cartesia's customers page lists ServiceNow, Together AI, Replicant, Thoughtly, Lorikeet, Fundamento, Bolna, Maven AGI, Forethought, Simple AI, 2X Solutions, Cresta, Rasa and others. Several rival voice-agent platforms in this index, including Vapi, Retell AI and LiveKit Agents, are themselves named as Cartesia customers and license Sonic and Ink as components of their own stacks.

| Fact | Value |
| --- | --- |
| Website | https://cartesia.ai |
| Pricing | A hybrid, two-meter model: a monthly subscription buys a pool of model credits (spent on Sonic TTS and Ink STT) plus a separate prepaid balance for Line voice-agent minutes, billed at $0.06/min ($0.014/min for telephony via a Cartesia phone number) regardless of tier. Free: $0/month (20K credits, $1 prepaid agent balance). Pro: $5/month (100K credits, $5 prepaid, commercial license, instant voice cloning). Startup: $49/month (1.25M credits, $49 prepaid, pro voice cloning, org accounts). Scale: $299/month (8M credits, $299 prepaid, priority support, higher concurrency). Enterprise: custom-quoted. Every tier includes unlimited workspace seats and voice slots, with cancel-anytime billing. |
| API | Yes |
| Best for | Engineering teams that want an open-source voice-agent SDK with real LLM choice, built on speech models several rival platforms in this category already license as a component. |
| Not for | Regulated enterprises that specifically need the voice-agent runtime itself (not just the speech models) fully self-hosted — Deepgram Voice Agent API offers that; Cartesia's Line layer does not. |

## Pricing

| Tier | Price |
| --- | --- |
| Free | $0 / month |
| Pro | $5 / month |
| Startup | $49 / month |
| Scale | $299 / month |
| Enterprise | Custom (contact sales) |

## Verdict

Cartesia's core credibility is that its speech models are already load-bearing infrastructure for part of this very category: Vapi, Retell AI and LiveKit Agents are all named Cartesia customers on its own site, using Sonic and/or Ink as a component of their own stacks. Line, the newer voice-agent layer built on top, is the most LLM-agnostic and the most genuinely open source of any entry here: an Apache 2.0 SDK that connects to 100+ providers via LiteLLM, backed by a real $0/month tier and flat $0.06/min agent pricing that doesn't escalate with plan. The tradeoff is deployment: unlike Deepgram Voice Agent API, only the speech models can be self-hosted, not Line's voice-agent runtime itself, so a buyer that needs the entire stack on its own infrastructure should look elsewhere in this category. A well-capitalized, fast-moving vendor ($191M raised, Series B in October 2025) still building out its independent review record.

## How it works

1. Start on Sonic (TTS) and Ink (STT) directly via API, or use Line for full voice-agent orchestration — via the no-code Agent Builder for prototyping, or the open-source Python SDK for production control.
2. In the SDK, connect any of 100+ LLM providers through LiteLLM (OpenAI, Anthropic, Google or open-weight models), define tools the agent can call, and configure multi-agent handoffs for specialized routing.
3. Deploy through Line's managed runtime — Cartesia hosts the compute, auto-scaling and audio pipeline; every call is recorded with transcripts and latency/system metrics available for review.
4. Billing draws from two separate pools per tier: a monthly credit allotment for Sonic/Ink usage, and a prepaid balance for Line voice-agent minutes at a flat $0.06/min ($0.014/min telephony) that doesn't change by tier.

## Who it's for

- Engineering teams that want an open-source voice-agent SDK (not a closed black box) with genuine LLM choice across 100+ providers, and are comfortable with Cartesia hosting the runtime rather than running it themselves.
- Teams prioritizing speech-model quality and latency first — Sonic and Ink are the core product several rival platforms in this category license as a component — who want the option to build the agent layer on top of the same models directly.
- Startups and solo builders who want a genuine $0/month starting point (20K TTS/STT credits plus a small prepaid agent balance) before committing spend, scaling into paid tiers that stay flat-rate on voice-agent minutes ($0.06/min) regardless of plan.

## Strengths and weaknesses

- ✓ A genuine $0/month free tier (20,000 model credits, a small prepaid agent balance) rather than only a time-limited trial, with flat $0.06/min agent pricing that doesn't change across paid tiers.
- ✓ Line's voice-agent SDK is open source (Apache 2.0, 103 GitHub stars as of this listing's verification) and connects to 100+ LLM providers via LiteLLM, rather than locking a buyer into one reasoning model.
- ✓ Well-capitalized and independently validated as production infrastructure: $191M raised across seed/A/B rounds (most recently a $100M Series B in October 2025 led by Kleiner Perkins with NVIDIA joining as a new strategic investor), and named as a speech-model supplier by several rival platforms in this index (Vapi, Retell AI, LiveKit Agents) on Cartesia's own customer page.
- ✓ Compliance built into the public API: SOC 2 Type II, PCI Level 1, GDPR, and HIPAA-eligible with a signed BAA available on enterprise contracts.
- ✗ Self-hosting only extends to the underlying Sonic and Ink speech models on an Enterprise contract. Line's voice-agent runtime itself always runs on Cartesia's managed infrastructure, unlike Deepgram Voice Agent API's fully self-hostable option in this same category.
- ✗ Pricing is a two-meter system (a monthly credit pool for Sonic/Ink usage, plus a separate prepaid balance specifically for Line agent minutes), which is less immediately legible than a single per-minute rate like Bland AI's.
- ⚠ Line's voice-agent runtime cannot be self-hosted — only the underlying Sonic/Ink models can, on an Enterprise contract.
- ⚠ No independent review-platform (G2/Capterra/Trustpilot) rating found as of this listing's verification.
- ⚠ Two separate billing meters (model credits and prepaid agent balance) rather than one blended per-minute rate.

## Key features

- **Sonic-3.6 (TTS) and Ink-2 (STT)** — Cartesia's own low-latency speech models, built on State Space Models rather than Transformers. Sonic-3.6 supports 40+ languages with laughter and emotion controls. Ink-2 handles streaming transcription.
- **Line SDK — open source, Apache 2.0** — A Python SDK (104 GitHub stars) for building voice agents, plus a separate no-code Agent Builder for prototyping without writing code.
- **100+ LLM providers via LiteLLM** — Bring OpenAI, Anthropic, Google Gemini or open-weight models to the agent layer. Line does not lock a buyer into one reasoning provider.
- **Multi-agent handoffs and built-in tool belt** — A main agent can transfer control to a specialized agent mid-call. Built-in tools include end_call, transfer_call, send_dtmf, voicemail detection, web_search, knowledge_base, and HTTP-server tools that turn JSON schemas into agent-callable APIs.
- **Enterprise self-hosting for the speech models only** — Sonic and Ink (not Line's voice-agent runtime) can be deployed into a customer's own GCP, AWS or on-premise environment on an Enterprise contract, for colocation, isolation or data-residency needs.
- **Native MCP server** — Cartesia publishes its own Model Context Protocol server (cartesia-mcp) so tools like Cursor, Claude Desktop and OpenAI Agents can call its TTS, STT, voice-cloning and pronunciation-dictionary APIs directly.

## Use cases

- **Building a voice agent with real LLM choice** — A developer uses the open-source Line SDK to wire in Anthropic Claude as the reasoning model while keeping Cartesia's Sonic/Ink for speech, avoiding lock-in to a single-vendor reasoning layer.
- **Prototyping without code** — A non-technical team member sketches an agent flow in the no-code Agent Builder before handing it to engineering to harden into the SDK for production.
- **Enterprise deployment with data-residency requirements** — A regulated enterprise self-hosts Sonic and Ink inside its own VPC under an Enterprise contract to keep audio processing inside infrastructure it controls, while still using Line's managed runtime for orchestration.

## Integrations

LiteLLM (100+ LLM providers, including OpenAI, Anthropic and Google Gemini, plus open-weight models) · Model Context Protocol — Cartesia publishes its own MCP server (cartesia-mcp) for Cursor, Claude Desktop and similar clients · Telephony via Cartesia-provisioned phone numbers ($0.014/min)

## Sources

- https://cartesia.ai/pricing
- https://cartesia.ai
- https://cartesia.ai/customers
- https://docs.cartesia.ai/line/introduction
- https://docs.cartesia.ai/self-hosted/introduction
- https://github.com/cartesia-ai/line
- https://research.contrary.com/company/cartesia