Today, we’re launching Cartesia Managed Agents, our end-to-end voice agent platform for enterprise conversations. It pushes the frontier of speech-to-speech naturalness and accuracy, and ranks #1 on independent voice agent benchmarks at launch.
We designed the harness and our voice models together, so listening, turn-taking, reasoning, tool-calling, and speech are tuned as one end-to-end system. Everything enterprise teams need to run agents in production comes on one platform: telephony, knowledge bases, tool integrations, and more. And we’ve tested it at scale — Managed Agents has already powered tens of millions of conversations in finance, retail, hospitality, and healthcare.
Setting a new Pareto frontier
Cascaded pipelines are the dominant paradigm for enterprise tasks, but without careful co-design they can lose out on naturalness and conversational dynamics like turn-taking and interruptions. Today’s speech-to-speech models sound natural, but they’re hard to control and less reliable for enterprise-grade tasks.
With Managed Agents, we co-designed our voice models and the harness to deliver the best conversational experience without sacrificing the intelligence and controllability that complex enterprise tasks require.
On Artificial Analysis’s Speech Agent Arena, where blind human raters compare agents on tasks like booking appointments and ordering takeout, Managed Agents ranks #1 of 30 agents with a preference Elo of 1115.
Speech Agent Arena: preference Elo
Blind human preference across conversational tasks like booking appointments and ordering takeout. Higher is better.
Benchmarks run with GPT-5.5 as the reasoning model.
On EVA-Bench, ServiceNow’s open benchmark for enterprise voice agents in CSM, ITSM, and HR, Managed Agents posts the highest scores for both accuracy (87%) and experience (91%), setting a new Pareto frontier.
EVA-Bench: accuracy vs. experience
First-attempt scores across 200+ enterprise scenarios. Every system in the shaded box scores lower on both. Higher is better.
Accuracy (EVA-A)
See every system's scores
| EVA-Bench: accuracy vs. experience | Accuracy (EVA-A) | Experience (EVA-X) |
|---|---|---|
| Managed Agents (GPT 5.5)Cartesia | 87.4% | 90.7% |
| Grok Voice 3.0xAI | 73.7% | 71.2% |
| Gemini 3.8 LiveGoogle | 45.9% | 88.6% |
| Gemini 3.8 Live Extended Thinking (Minimal)Google | 50.1% | 79.2% |
| Gemini 3.8 Live Extended Thinking (High)Google | 54.8% | 73.7% |
| ElevenLabs Agents (Claude Haiku 4.5)Scribe v2.2 Realtime + Claude Haiku 4.5 + Eleven Flash v2 | 41.0% | 81.7% |
| GPT Realtime 2.1OpenAI | 59.0% | 61.6% |
| Grok Voice 2.0xAI | 51.7% | 64.1% |
| Grok Voice 1.0xAI | 58.8% | 56.9% |
| GPT Realtime 1.5OpenAI | 46.7% | 56.6% |
| GPT Realtime 2OpenAI | 50.4% | 39.0% |
| Gemini 3.1 Flash LiveGoogle | 29.2% | 58.9% |
| ElevenLabs Agents (GPT-5.4)Scribe v2.2 Realtime + GPT-5.4 + Eleven Flash v2 | 67.2% | 13.0% |
| Universal 3.5 Pro + GPT-5.4 + Sonic 3.5Cascade | 53.1% | 24.8% |
| GPT RealtimeOpenAI | 40.7% | 36.5% |
| Parakeet 1.1B CTC en-US + Gemma-4-26B + KokoroCascade | 33.1% | 39.5% |
| Flux General Multi + GPT-5.4 + Sonic 3.5Cascade | 43.1% | 17.6% |
| GPT Realtime MiniOpenAI | 16.3% | 40.6% |
| Gemini 3.6 Flash + Gemini 3.1 Flash TTS PreviewGoogle | 56.2% | 0.0% |
| ElevenLabs Agents (Gemini 3 Flash)Scribe v2.2 Realtime + Gemini 3 Flash + TTS Conversational v3 | 49.0% | 2.4% |
| GPT Realtime 2.1 MiniOpenAI | 21.7% | 29.2% |
| Nova 3 + GPT-5.4-mini + Aura 2Cascade | 39.5% | 9.6% |
| Ink 2 + Haiku 4.5 + Sonic 3.5Cascade | 34.8% | 6.6% |
| UltravoxHybrid | 27.0% | 2.9% |
Benchmarks run with GPT-5.5 as the reasoning model.
On EVA-Bench, ServiceNow’s open-source benchmark for enterprise voice agents, Cartesia Managed Agents posted one of the strongest results we’ve seen to date. Across 200+ realistic CSM, ITSM and HR scenarios, it stayed accurate on the details while keeping conversations natural, which is the balance enterprises look for before putting AI on the phone with their customers. It also reflects a broader pattern we’re seeing in enterprise voice: harnesses and models designed together.
When callers interrupt, pause, or change their minds, the agent handles it naturally: in our internal evaluations, callers reported no turn-taking issues in 85% of conversations with Managed Agents, compared with 67% for other agents.
Designed for complex calls
Cartesia Managed Agents provides all the tools needed to build frontier voice agents in one end-to-end platform. Claims intake, patient scheduling, retail returns, IT helpdesk: enterprise calls rarely stay simple. One conversation can mean verifying a caller, checking three systems, and quoting policy, all without missing a beat.
- Multilingual: Build agents in 44 languages. Agents can switch between English, Spanish, Hindi, French, and Japanese mid-conversation, with all 44 coming soon.
- Tool calls: Integrate your backend APIs, CRMs, and systems of record, so your agent can look up records and take action mid-call.
- Knowledge bases: Get precise answers from hundreds of documents. The agent’s context is prefilled dynamically, so it responds without pausing to search.
- Outbound campaigns: Schedule thousands of calls with outbound batches, each personalized to the person you’re calling.