Cartesia Managed Agents Claims #1 on Voice Agent Benchmarks: What the Live Leaderboard Shows and What It Does Not

Cartesia launched Managed Agents on October 29, an end-to-end voice agent platform for enterprise calls. The launch post leads with benchmark rank: "#1 of 30 agents" on Artificial Analysis's Speech Agent Arena, and the highest accuracy and experience scores on ServiceNow's EVA-Bench. I opened the live arena page and the EVA README to see how those claims read at the source. The short version: the arena shows Cartesia first, but with a confidence range that includes second place, and the arena's own task success column does not put it first.

I did not use Managed Agents, run EVA-Bench, or read the arena methodology page or the EVA paper. Product claims are Cartesia's.

What the launch post says

• Managed Agents is "our end-to-end voice agent platform for enterprise conversations", with the harness and voice models designed together so listening, turn-taking, reasoning, tool calling and speech are "tuned as one end-to-end system".
• It "has already powered tens of millions of conversations in finance, retail, hospitality, and healthcare". I have no source for that beyond the post.
• Speech Agent Arena: "ranks #1 of 30 agents with a preference Elo of 1115", from blind human raters on tasks like booking appointments and ordering takeout. The post's chart lists, among others, Gemini 3.1 Flash Live Minimal at 1099, Gemini 3.8 Live at 1079, Grok Voice 3.0 at 1064 and GPT-Live-1 (Sol, low) at 1041. Footnote: "Benchmarks run with GPT-5.5 as the reasoning model."
• EVA-Bench: "the highest scores for both accuracy (87%) and experience (91%)". The post's table gives Managed Agents (GPT 5.5) at 87.4% accuracy (EVA-A) and 90.7% experience (EVA-X), against Grok Voice 3.0 at 73.7% and 71.2%, and Gemini 3.8 Live at 45.9% and 88.6%.
• A quote from a ServiceNow product management director in the post calls it "one of the strongest results we've seen to date".
• An internal evaluation: "callers reported no turn-taking issues in 85% of conversations with Managed Agents, compared with 67% for other agents."
• Features: 44 languages (the post says switching mid-conversation works among English, Spanish, Hindi, French and Japanese, "with all 44 coming soon"), tool calls, knowledge bases, outbound campaigns, telephony.

What the live arena page showed

I opened artificialanalysis.ai/speech-to-speech/arena on October 31, 2026. Its description: preference across blind, live voice conversations on topics like booking a dental appointment and ordering takeout. Overall Preference view:

Rank	Range	System	Elo	Samples	Task success rate
1	1 to 2	Cartesia Managed Agents	1111 ±29	685	91.7%
2	1 to 3	Gemini 3.1 Flash Live Minimal	1089 ±25	990	74.6%
3	2 to 5	Gemini 3.8 Live	1074 ±27	606	93.2%
5	3 to 6	Grok Voice 3.0	1059 ±28	668	94.8%

Things I notice, which are my reading of the page and not claims by Artificial Analysis:

• The page shows Elo 1111, the post says 1115. The page ranks it 1 with a range of 1 to 2, so a tie for first is inside the interval. The numbers on a live board move as samples arrive, so a small gap is not surprising, but I cannot say when the post's numbers were taken beyond "October 2026".
• On the page's task success column, Cartesia (91.7%) is below Gemini 3.8 Live (93.2%) and Grok Voice 3.0 (94.8%), though confidence intervals overlap (Cartesia 88.3% to 94.3%).
• The page lists 30 systems, which matches "of 30". It also marks some entries as "Default Cascaded System", meaning a provider's standard pipeline rather than one tuned by Artificial Analysis.
• The post's footnote says the benchmarks ran with GPT-5.5 as the reasoning model, and the EVA row is labeled "Managed Agents (GPT 5.5)". The score is for the platform with that model underneath, not for Cartesia's voice models alone. I did not see on the arena page which reasoning model each entry uses.

What the EVA README says

ServiceNow's EVA repo describes an open-source framework that scores multi-turn spoken conversations on EVA-A (accuracy: did the agent complete the task correctly and faithfully) and EVA-X (experience: natural, concise, appropriate for speech). It uses a bot-to-bot setup with no human listeners, and it includes 213 enterprise scenarios across 3 domains, plus results for 12 systems and a perturbation suite for noise, accents and degraded connections.

Comparing with the post:

• The post says 200+ scenarios. The README says 213. Consistent.
• The README says results for 12 systems. The post's table lists 24. I did not read the EVA leaderboard, so I cannot tell whether the post's other rows come from ServiceNow's published results or from runs Cartesia or others added, or whether the Managed Agents row is on ServiceNow's board. I would check before treating the table as ServiceNow's.
• The README describes a perturbation suite (noise, accents, connection degradation). The post does not say whether the results include it.

What I would check

These are my untested suggestions.

1. Open the arena page yourself and read the range and task success columns, not only the rank.

2. Check the EVA leaderboard for a Managed Agents row, and which reasoning model and settings it used.

3. Ask what is held constant. The same reasoning model for every row would make the comparison cleaner, and the post says GPT-5.5 was used for its runs.

4. Run your own calls. Scripted booking tasks are not your policy lookups or your accents. Test with your real scenarios, noise and interruptions.

5. Separate voice quality from task accuracy. The post's EVA numbers cover both, but they come from different metrics.

6. Treat the "tens of millions of conversations" and the 85% vs 67% turn-taking figure as unverified until Cartesia shows how they were measured.

7. Check language coverage for the ones you need. The post says all 44 are "coming soon".

What I would not conclude

• That Cartesia is the best voice agent platform. The arena range includes second place, and its task success rate is below others on the live page.
• That the benchmark gap shows up on your workload.
• That the EVA table is ServiceNow's. I did not verify that.
• That the live page and the post disagree in a meaningful way. The Elo difference is small and boards update.

Frequently asked questions

What is Cartesia Managed Agents?

Per Cartesia, an end-to-end voice agent platform for enterprise conversations with telephony, knowledge bases, tool integrations and outbound campaigns.

Is it really number one on the Speech Agent Arena?

On the live page I read on October 31, it is ranked 1 with a range of 1 to 2 and Elo 1111. The post says 1115 and "#1 of 30".

What is EVA-Bench?

An open-source ServiceNow framework that scores spoken agent conversations on accuracy (EVA-A) and experience (EVA-X), per its README.

How many scenarios does EVA have?

The README says 213 enterprise scenarios across 3 domains. The post says 200+.

Which reasoning model was used?

The post's footnote says GPT-5.5, and its EVA row is labeled Managed Agents (GPT 5.5).

Does it support many languages?

The post says 44 languages, with mid-conversation switching among five and the rest "coming soon".

Did you try the product?

No. I read the launch post, the arena page and the EVA README.

Sources

All read on October 31, 2026.

• Introducing Managed Agents (Cartesia, October 29, 2026)
• Speech Agent Arena Leaderboard (Artificial Analysis)
• ServiceNow/eva README

Not read: the arena methodology and overview pages, the EVA leaderboard and paper, and Cartesia's product documentation.