Introduction

Cartesia just launched Managed Agents, its end-to-end platform for building voice agents that answer phones and talk to customers. It's closed and hosted: you write the instructions, pick an LLM and a voice, and Cartesia runs the listening, turn-taking, interruptions, tool calls and speech around it. You can build and talk to one today in the Cartesia Playground or create agents through the API.

The headline number from the October 29 launch post: Managed Agents ranks #1 of 30 agents on Artificial Analysis's Speech Agent Arena with a preference Elo of 1115. Call time costs $0.03 per minute, and the free plan includes $1 of agent credit, roughly 33 minutes, to try it.

What Shipped

From Cartesia's launch post and Managed Agents docs:

  • One system, not a pile of parts: Cartesia says it tuned its voice models and the agent "harness" together, so listening, turn-taking, reasoning, tool calls and speech behave as one pipeline instead of three separate services stitched together.
  • Bring a model, no keys: pick an LLM from OpenAI, Anthropic or Google inside Cartesia. You don't need your own provider account or API key, and you can set automatic fallback models.
  • Tools mid-call: built-in system tools (like ending or transferring a call), webhook tools that hit your server, and client tools that run in your own app over the WebSocket API.
  • Knowledge bases: answers from hundreds of documents, with context prefilled so the agent doesn't pause to search.
  • Phones and outbound batches: Cartesia phone numbers, your own SIP trunk, and scheduled outbound campaigns personalized per caller.
  • Languages: agents can be built in 44 languages. Switching languages mid-conversation works now for English, Spanish, Hindi, French and Japanese, with the rest "coming soon."
  • Evaluations and versions: built-in call-quality metrics, plus a version history every time you change the model or config.

The Benchmarks

Cartesia is leaning hard on two third-party leaderboards, both run with GPT-5.5 as the reasoning model:

  • Speech Agent Arena (Artificial Analysis): blind human raters compare agents on tasks like booking appointments and ordering takeout. Managed Agents scored 1115, ahead of Gemini 3.1 Flash Live Minimal at 1099, Gemini 3.8 Live at 1079 and Grok Voice 3.0 at 1064. ElevenLabs' Eleven Realtime v1 Max sat at 1003.
  • EVA-Bench (ServiceNow): an open benchmark of 200+ enterprise support, IT and HR scenarios. Managed Agents posted 87.4% accuracy and 90.7% experience, the top score on both. Next best on accuracy was Grok Voice 3.0 at 73.7%, and next best on experience was Gemini 3.8 Live at 88.6%.

Cartesia also says callers reported no turn-taking problems in 85% of conversations with Managed Agents versus 67% for other agents, but that one is from its own internal evaluation, so treat it as a vendor claim. It also says the platform has already handled tens of millions of conversations before today's public launch.

Pricing

From Cartesia's pricing page:

  • Call time: $0.03 per minute on every plan.
  • Telephony: $0.014 per minute extra when you use a Cartesia-provided phone number.
  • LLM usage: billed separately on top of the per-minute rate.
  • Free plan: $1 of agent credit (about 33 minutes) and one Cartesia phone number, with 8 concurrent calls.

How to Try It Today

  1. Open the agents Playground and create an agent.
  2. Write the instructions and a welcome message, then pick a voice and an LLM.
  3. Talk to it in the browser first. When it behaves, attach a Cartesia number or your SIP trunk.

If you'd rather script it, the same agents live in the API, and this is how you pin the LLM in the agent config:

{
  "config": {
    "model": {
      "id": "gpt-5.4-mini",
      "temperature": 0.3,
      "max_output_tokens": 1024
    }
  }
}

Call GET /v1/agents/models first to see which model IDs your account has, along with each one's latency and token price.

Pro Tip

Before you trust any leaderboard, steal the arena's own test: give your agent one booking task and have a friend call it and interrupt halfway, change their mind, and spell out a phone number. Then switch the LLM and run the same call again. Every model change creates a new version, so you can roll back if the "smarter" model turns out to be slower on the phone.

Why It Matters for Audio Creators

Voice agents are where TTS and speech recognition actually earn money right now, and this week has been crowded: Phonic-1 went the single speech-to-speech model route, while Cartesia is betting that a carefully tuned pipeline of separate models can beat it on both accuracy and naturalness. If you build talking characters, hotlines for a project, or interactive audio experiences, a free 33 minutes is enough to find out whether that bet holds for your use case.

Conclusion

Cartesia Managed Agents is a closed, hosted voice agent platform built on its own speech models, ranked #1 on Artificial Analysis's Speech Agent Arena at launch, and priced at $0.03 per minute plus LLM and phone costs. Spin one up in the Playground, interrupt it on purpose, and see whether it keeps up.

—Anabel ♡