#1 on EVA-BenchEVA-Bench is ServiceNow's independent benchmark for voice agents in enterprise workflows like customer support. It scores each agent on how accurately it completes the task and on the caller's experience.
#1 on Speech Agent ArenaSpeech Agent Arena is Artificial Analysis' independent benchmark for grading speech to speech models on enterprise tasks using human evaluators.
Callers prefer Cartesia Managed Agents to the Gemini Live API
Cartesia hosts the agent and runs every call, including its phone number, transfers, and hang-ups, so your server doesn't have to.
Caller preference
Artificial Analysis Speech Agent Arena Elo, October 2026
1116
Managed Agents
1098
Gemini 3.1 Flash Live Minimal
Phone numbers
Each vendor's documentation, checked October 2026
Built in
$0.014/min
Partner
Through Voximplant, Twilio, or similar
Hear why callers prefer Cartesia to Gemini Live
Weighing the Gemini Live API? Pick a job, talk to a live Cartesia agent in your browser, then clone it to start building.
Caller preference
Callers prefer Cartesia Managed Agents to every Gemini Live model
- Artificial Analysis puts people in blind, live voice conversations with two agents on tasks like booking an appointment, then records which one they prefer. Elo turns those choices into one rating: the higher it is, the more often people preferred that agent.
- Cartesia Managed Agents ranks first of 30 agents, 18 points above Gemini 3.1 Flash Live Minimal, Google's highest-rated model.
Speech Agent Arena: preference Elo
Higher is better.
Task success
Cartesia Managed Agents completes more tasks than Gemini 3.1 Flash Live and Gemini 3.8 Live Extended Thinking
- In the same conversations, Artificial Analysis checks whether the agent made the tool call that finishes the caller's request, such as booking the slot or placing the order. Task success rate is the share of conversations where it did.
- Cartesia Managed Agents completes 91.7% of tasks, against 74.6% for Gemini 3.1 Flash Live Minimal.
Task success rate
Share of conversations that end with the task done. Higher is better.
What each platform includes
Compare how each platform connects to your systems, handles calls, and bills for usage. Check what is included before choosing a plan.
Caller preference
People hold blind, live conversations with two agents and pick the one they would rather keep talking to.
Cartesia Managed Agents1116 Elo#1 of 30 agentsGemini Live API1098 EloGemini 3.1 Flash Live Minimal, Google's highest ratedOctober 2026
Models
Cartesia Managed AgentsInk 2 listens, Sonic 3.6 speaks, and you choose the LLM from OpenAI, Anthropic, and Google Gemini models.Gemini Live APIGemini Live models only, such as Gemini 3.8 Live.Phone calls
Cartesia Managed AgentsCartesia-provided phone numbers at $0.014/min.Gemini Live APINo Google phone numbers. Calls connect through a partner such as Voximplant or Twilio.Call actions
Cartesia Managed AgentsBuilt-in end call, transfer, keypad tones, and voicemail detection, plus webhook and client tools.Gemini Live APIFunction calling. Your code runs each function, including transfers and hang-ups.Setup
Cartesia Managed AgentsConfigure and test agents in the Cartesia playground or through the API. Cartesia hosts each call.Gemini Live APIYour server or app opens a WebSocket and runs the session.
Appointment calls booked in your own calendar
Pick a voice and an AI model, write instructions, and connect your booking system.
Connect your booking system
The agent checks open times and books the slot during the call.
Let Cartesia handle the voice
Cartesia handles listening, turn-taking, and speech. Callers can interrupt, and the agent follows.
Move one Gemini Live workflow to Cartesia
Recreate one agent by hand, test it with your own calls, then move the number.
Follow the agent configuration guideRecreate the session as an agent
The system instruction carries over as text, and each function your code ran becomes a tool Cartesia calls during the call.
Test the agent, then move your production calls
Call the agent from your browser with real booking requests and score every call before you move the number.
Frequently asked questions
What does the Speech Agent Arena measure?
Artificial Analysis's Speech Agent Arena puts people in blind, live conversations with two voice agents on tasks like booking appointments and ordering takeout, and records which agent they prefer. It also scores whether each conversation ends with the tool call that completes the task. The results on this page are from the October 2026 release, with GPT-5.5 as Managed Agents' LLM.
Can Managed Agents use a Gemini model?
Yes. Managed Agents' model catalog includes Google Gemini models, along with OpenAI and Anthropic models. Cartesia manages the integration, so you don't need a Google account or API key. Cartesia's Ink 2 listens and Sonic 3.6 speaks, whichever model you choose.
How is Managed Agents different from the Gemini Live API?
The Gemini Live API connects your server or app to Gemini's audio models over a WebSocket. Your code handles phone calls, transfers, and the actions the model asks for. Managed Agents is a hosted agent: you configure it in the Cartesia playground or through the API, and Cartesia runs the call, its phone number, and built-in actions such as transfers and hang-ups.
Can I import my Gemini Live setup?
There is no automatic import. Recreate one agent's instructions and actions in your Cartesia dashboard, test it with your own calls, then move the number.
How much does Managed Agents cost?
Managed Agents costs $0.03 per minute plus the LLM. A Cartesia-provided phone number adds $0.014 per minute. See the pricing page for plan terms.
Sources
Gemini Live API pricing and features checked on .
More comparisons
Get started today
Talk to an expert.
Connect with a member of our team and learn how Cartesia can help you build world-class voice experiences.
Start building.
Access our models via API and bring a voice agent into production in minutes.