Cartesia's Voice Agent Tops the Chart: 3 Cents per Minute, Ahead of Google and OpenAI

Cartesia has released Managed Agents, an end-to-end voice agent platform for enterprises. On launch day, it took first place on two independent leaderboards at the same time.

Let's look at the data first.

On Artificial Analysis's Speech Agent Arena, which uses blind human ratings for everyday conversation tasks like booking appointments and ordering takeout, Managed Agents ranks first with a preference Elo of 1115. Behind it are Google's Gemini 3.1 Flash Live Minimal (1099), Gemini 3.8 Live (1079), xAI's Grok Voice 3.0 (1064), and OpenAI's GPT-Live-1 (1041 and 1037). ElevenLabs's Realtime v1 Max is at 1003.

The other is ServiceNow's EVA-Bench, covering more than 200 enterprise scenarios such as customer service, IT helpdesk, and HR. Managed Agents has an accuracy of 87.4% and an experience score of 90.7%, both the highest. The second-place Grok Voice 3.0 has an accuracy of 73.7% and an experience score of 71.2%. The gap is not small.

Cartesia's explanation is that they designed the voice model and the harness together. Listening, turn-taking, reasoning, tool calling, and voice output are tuned as one system, rather than stitching together ASR, LLM, and TTS as three separate segments.

This claim is worth noting. In the past, the mainstream approach for enterprise voice agents was a cascaded pipeline, optimizing each segment separately, but information is lost between segments, and common actions in real conversations—interruptions, pauses, changing one's mind—are often handled poorly. End-to-end voice models are more natural but less controllable, and complex tasks can easily go off track. Cartesia is betting on a third path.

At the product level, Managed Agents packages what enterprises need to run voice agents into one platform:

• Telephony, supporting Cartesia, Twilio, or SIP numbers, for both inbound and outbound calls. Outbound batch processing can queue up to 5,000 numbers at a time.
• Tool calling, connecting to backend APIs, CRMs, and systems of record, to look up data and perform actions during calls.
• Knowledge base, retrieving answers from hundreds of documents, with dynamic context pre-filling, so it doesn't have to stop to search.
• Multilingual, 44 languages, with mid-call switching between English, Spanish, Hindi, French, Japanese, and others.
• Testing and archiving, with direct test calls in the browser, custom metric scoring, and recording, transcription, and summary retention for every call.

On compliance, there is HIPAA, SOC 2 Type 2, and GDPR. Cartesia says the system has already run tens of millions of conversations in finance, retail, hospitality, and healthcare.

Pricing is another key point. 3 cents per minute, with a 50% discount during the launch period. If you use Cartesia's own phone number, it's an additional 1.4 cents per minute.

Some people commented under the tweet that 3 cents a minute could make after-hours customer support affordable. Others said the first thing to be disrupted may not be customer service, but those human hotlines where people are still waiting in queue.

There's another data point that hasn't been mentioned much. In Cartesia's internal evaluation, 85% of conversations had no turn-taking issues, compared with 67% for other agents. This number comes from the vendor's own testing, not a third party, so it should be discounted. But the problem it points to is real: the hardest part of a voice agent is often not whether it answers correctly, but whether it can stop when it should stop and pick up when it should pick up.

The EVA-Bench results also provide a reference. Gemini 3.8 Live's experience score is 88.6%, which is not low, but its accuracy is only 45.9%. ElevenLabs's cascaded solution has an accuracy of 67.2%, but its experience score drops to 13.0%. Raising both dimensions at the same time is harder than raising just one. Cartesia has put both numbers in the upper right corner this time.

Of course, leaderboards have their limits. Speech Agent Arena's tasks lean toward everyday scenarios—booking appointments, ordering takeout—and are not the same as truly complex enterprise calls. EVA-Bench is closer to enterprise scenarios, but whether 200-plus scenarios can cover the real long tail still needs time to verify.

What is truly certain is the price. 3 cents per minute means a 10-minute call costs 30 cents. Once this number holds, the logic for deploying voice agents will change. The question is no longer whether to use them, but which scenarios haven't used them yet.

Cartesia's documentation includes an FAQ: Is it legal for AI to make calls? The answer is to comply with the laws in your region, such as the TCPA in the U.S., obtain consent first, respect opt-outs, and check industry and jurisdiction rules. Putting this sentence on the launch page is quite clear-headed.

Technology moves faster than rules, and this time is no exception.