Cartesia vs Deepgram (2026)

Managed Agents take webhook tools, where Cartesia sends the HTTPS request to an endpoint the customer exposes, client tools that run in the customer's app over the WebSocket, and system tools for ending a call, sending DTMF and transferring to a phone number, up to 32 tools per agent, so an agent can look up an order or trigger an action during a call. Plugins for LiveKit, Pipecat and other voice stacks distribute Cartesia's models and are not agent integrations.

SourceCartesia, docs.cartesia.ai (Tools, Webhook tools, Integrations)read 2026-09-27

Voice Agent function calling lets the customer define custom tools as functions in Settings. Each function runs client side in the customer's application, or server side, where Deepgram calls a web endpoint the customer provides. During a live call, an agent can book appointments, send emails or update records in the customer's CRM, ERP or internal APIs.

SourceDeepgram, developers.deepgram.com (Function Calling)read 2026-09-27

Cartesia runs a fixed agent loop. A Managed Agent combines Ink-2 listening, the chosen LLM and Sonic speech with turn taking, interruptions and tool calls, and each configuration change is saved as a version. No workflow graph, branching model or hand off between agents is described for Managed Agents, and the older Line SDK is being migrated away from.

SourceCartesia, docs.cartesia.ai (Managed Agents, Agent configuration, Migrate from the Line SDK)read 2026-09-27

Each Voice Agent session is a fixed pipeline that Deepgram runs end to end. It chains listen, think and speak, with an ordered LLM fallback chain per request and mid call Update messages for the prompt and providers. The customer configures the pieces, not the control flow. In the multi agent architecture guide, a CallOrchestrator sequences qualifier, advisor and closer agents in a reference repository the customer runs, so that orchestration is the customer's own code.

SourceDeepgram, developers.deepgram.com (LLM Models, Build a Multi-Agent Architecture)read 2026-09-27

A call batch stores up to 5,000 recipients, and Cartesia dials them in the background, at a scheduled_at time when one is set, up to a concurrency limit, with retries for failed calls. Agents also take inbound and outbound calls on Cartesia or Twilio numbers and SIP trunks, and stream over a WebSocket from web and mobile apps.

SourceCartesia, docs.cartesia.ai (Batch calling, Phone numbers, WebSocket API)read 2026-09-27

The Voice Agent works on phone calls through inbound and outbound telephony builds for Twilio and Amazon Connect, and in web pages through the Browser Agent SDK and an embeddable widget. Every session starts when a caller dials in, or when the customer's own server opens the socket or places the call. The outbound reference build takes a POST from the customer's CRM or CLI. Deepgram has no event, schedule or webhook of its own that wakes the agent.

SourceDeepgram, developers.deepgram.com (Build an Outbound Telephony Agent, Browser Agent SDK)read 2026-09-27

Documents uploaded through the Playground or the documents API, singly or in bulk up to 100 a request, are indexed automatically on upload and organized into nested folders with metadata for filtering. An agent's queries search only the folders attached to it, and documents can be updated or deleted through the API.

SourceCartesia, docs.cartesia.ai (Knowledge Base, Create Document API)read 2026-09-27

The products cover speech to text, text to speech, audio and text intelligence, and the Voice Agent, and none of them ingests, indexes or retrieves the customer's documents. Context reaches the agent through the prompt, loaded history and function results the customer supplies. Keyterm prompting tunes transcription vocabulary.

SourceDeepgram, developers.deepgram.com (llms.txt, Maintaining Context)read 2026-09-27

Within a call the agent keeps the conversation, and dynamic variables set at the start or changed by tools are read by the agent each time it responds. Nothing carries into the next call except what the customer passes in as variables, and no memory layer with a scope and lifetime is described.

SourceCartesia, docs.cartesia.ai (Agent configuration, Dynamic variables)read 2026-09-27

Conversation state lives inside a session. The agent keeps the prompt, turns, injected messages and function results as working memory for the call. History (on by default) lets a new session load prior turns and function calls through agent.context.messages. The customer's application stores and supplies those prior turns. Deepgram names no memory layer of its own that lasts beyond a call.

SourceDeepgram, developers.deepgram.com (Maintaining Context, History)read 2026-09-27

The transfer system tool hands the caller to a phone number, and turn settings end or check in on silent calls. The configuration's Guardrails are rules written into the agent's instructions, and no approval step holds a tool call for a person.

SourceCartesia, docs.cartesia.ai (Agent configuration, System tools)read 2026-09-27

The customer can hold back agent actions, but there is no approval step. Setting defer_until_eot on a function holds any call whose side effect cannot be undone (ending a call, spending money, sending a message) until speech to text confirms the caller has finished the turn. Read only functions still dispatch early. Deepgram advises builders to set it on any irreversible action. No person approves an action inside Deepgram.

SourceDeepgram, developers.deepgram.com (Function Calling: Irreversible Actions and Turn Confirmation)read 2026-09-27

Enterprise organizations get self serve SSO over SAML or OIDC, configured by an organization admin, and Admin and Member roles, where only admins manage members, invitations and settings. Cartesia states it is SOC 2 Type II certified, a Vanta trust center sits at trust.cartesia.ai, and Zero Data Retention is available on Enterprise.

SourceCartesia, docs.cartesia.ai (Set up SSO, Set up an organization) and cartesia.ai/pricingread 2026-09-27

Access runs on account permissions plus owner, admin and member project roles, each implying a listed set of scopes. API keys are created per project and scoped to those permissions. Deepgram has SOC 2 Type 1 and Type 2 reports from an independent auditor, with certificates on request. It is a HIPAA business associate and signs a BAA for Enterprise customers. It also states GDPR, CCPA and PCI compliance, reviewed yearly.

SourceDeepgram, developers.deepgram.com (Working With Roles & API Scopes, Data Privacy Compliance) and deepgram.com/pricingread 2026-09-27

Every call has a transcript with turn level timestamps and call logs with the agent's logging statements, custom events can be recorded, runtime logs are retrievable through the API, and webhooks deliver call and turn events carrying each turn's tool calls, variable updates and speech latency, plus a post call analysis.

SourceCartesia, docs.cartesia.ai (Observability, Get Call Runtime Logs)read 2026-09-27

The Console shows usage and up to 90 days of request logs, and a Usage API exports them to tools such as Grafana or Datadog. For the Voice Agent, the dashboard gives no per session, turn by turn observability, and there is no separate logging API. The customer taps the WebSocket and stores the transcript, function call, latency and error events it emits. The per session trace is the customer's to keep.

SourceDeepgram, developers.deepgram.com (Session Observability, Logs & Usage Data)read 2026-09-27

Enterprise customers can use dedicated regional deployments in the United States, European Union, United Kingdom, India and Australia that keep inference traffic in region. The same models and agents run cloud, on premise and on device, and on premise and VPC deployment are arranged through sales.

SourceCartesia, docs.cartesia.ai (Regional endpoints) and cartesia.ai (home, pricing FAQ)read 2026-09-27

Regional endpoints at api.eu.deepgram.com (EU, never routed outside it) and api.au.deepgram.com (Australian infrastructure for storage and inference) serve speech to text, text to speech and the Voice Agent with the same keys. Premium customers can self host in cloud instances they requisition, such as AWS or GCP, or in their own data center. A third party LLM in the think step runs on that provider's infrastructure.

SourceDeepgram, developers.deepgram.com (Regional Endpoints, Deployment Options, Data Privacy Compliance)read 2026-09-27

The agents API lists public, Cartesia provided agent templates to help a customer get started, but no template or the job it does is named, and there is no whole agent a customer adopts. The voice library, multilingual voices and pinned model snapshots are model assets, not agents.

SourceCartesia, docs.cartesia.ai (List Templates API reference)read 2026-09-27

There are no packaged agents for customers to adopt. The Voice Agent template apps are one starter ported to twelve languages and frameworks on GitHub, offered as sample code. Reusable Agent Configurations store the customer's own agent blocks. Nova-3 Medical and Pharma are industry tuned speech models, not agents.

SourceDeepgram, developers.deepgram.com (Template Apps, Reusable Agent Configurations)read 2026-09-27

The customer chooses the model. Each agent's config sets the LLM by ID from the list GET /v1/agents/models returns, with provider, average latency and token prices shown, plus temperature and output limits, and changing it creates a new version so candidates can be compared. Cartesia manages the provider accounts, and speech in and out uses Cartesia's own Ink and Sonic models.

SourceCartesia, docs.cartesia.ai (LLMs, Agent configuration)read 2026-09-27

Each Voice Agent's Settings name the think provider and model. The customer picks from OpenAI, Anthropic, Google and NVIDIA models Deepgram manages, or points to its own endpoint for Groq, Amazon Bedrock or any OpenAI compatible API. An ordered array of providers acts as a per request fallback chain that can mix providers. Listen and speak models are chosen the same way, including BYO TTS.

SourceDeepgram, developers.deepgram.com (LLM Models)read 2026-09-27

A versioned REST API, with 214 reference pages and published OpenAPI and AsyncAPI specs, covers text to speech, speech to text, voices and agents, including agents, tools, knowledge base documents, call batches, metrics and webhooks. Official JavaScript and TypeScript and Python SDKs, a CLI, and an MCP server that runs TTS, STT, voices and pronunciation dictionaries sit beside it.

SourceCartesia, docs.cartesia.ai (llms.txt, Client Libraries, MCP)read 2026-09-27

REST and WebSocket APIs cover speech to text, text to speech, text intelligence and the Voice Agent. A management API handles projects, keys and usage, and an Agent Configuration API sits beside it. Python, JavaScript, Go and .NET SDKs come with a feature matrix, alongside a Browser Agent SDK and the dg CLI, whose built in MCP server proxies the developer API's tools.

SourceDeepgram, developers.deepgram.com (llms.txt, Reusable Agent Configurations, MCP Server)read 2026-09-27

Completed calls are scored with built in metrics, such as call success and speech latency, and with custom LLM as a judge metrics the customer writes and assigns to an agent, with results exportable, so output quality can be tracked over time. There is no fixture or dataset testing before production, quality gate or tuning loop.

SourceCartesia, docs.cartesia.ai (Metrics, Results)read 2026-09-27

There is no evaluation harness, scored test cases or quality gate for voice agents. Testing tools cover integrations, such as a streaming starter kit and a SageMaker endpoint check. Reusable Agent Configurations list A/B testing of voices or prompts as a use case, but the customer measures conversion or CSAT on its own. Custom model training improves transcription.

SourceDeepgram, developers.deepgram.com (llms.txt, Reusable Agent Configurations)read 2026-09-27

The DTMF system tool presses phone keys, which is not browser or desktop control, and the browser examples stream voice into a web page, the product running in a browser. There is no browser, desktop or computer control.

SourceCartesia, docs.cartesia.ai (llms.txt, System tools)read 2026-09-27

The Browser Agent SDK embeds a voice agent in a web page, so the product runs in a browser without operating one. Telephony is a voice channel. Deepgram names no browser, desktop or computer control for its agents.

SourceDeepgram, developers.deepgram.com (llms.txt, Browser Agent SDK)read 2026-09-27