Cartesia launched Managed Agents on October 29, 2026, an end-to-end platform for enterprise voice agents. In its launch post, the company says the product "ranks #1 on independent voice agent benchmarks at launch," citing Artificial Analysis's Speech Agent Arena (preference Elo 1115, first of 30 agents) and ServiceNow's EVA-Bench (87 percent accuracy, 91 percent experience). Launch pricing is 50 percent off, "just $0.03 a minute."
Benchmark leadership is easy to claim and hard to compare. This article reads the numbers, notes which part a third party has confirmed, and lists what to test before you move a phone line to a new voice stack.
TL;DR: claimed vs. verified #
| Claim | Source | Status |
|---|---|---|
| #1 of 30 on Speech Agent Arena, Elo 1115 | Cartesia post | Shown in vendor's chart; we did not confirm on the Artificial Analysis leaderboard |
| Highest accuracy (87.4%) and experience (90.7%) on EVA-Bench | Cartesia's chart | Shown in vendor's chart; quote from a ServiceNow director |
| Callers reported no turn-taking issues in 85% of conversations vs 67% for other agents | Cartesia internal evaluations | Vendor-run, unaudited |
| Tens of millions of conversations in finance, retail, hospitality and healthcare | Cartesia | Vendor claim |
| $0.03 a minute, 50% off launch price | Cartesia | Vendor pricing; limited-time |
| Benchmarks run with GPT-5.5 as the reasoning model | Cartesia's chart footnotes | Disclosed |
What is Managed Agents? #
Cartesia describes it as a vertically integrated platform. In its docs, Managed Agents "handles listening, turn-taking, interruptions, tool calls, and knowledge retrieval around the LLM of your choice." You configure instructions, welcome message, voice and LLM in a playground or through the API, get an agent ID, and connect browser, app or server audio over WebSocket, or a phone number through telephony. Three kinds of tools are supported: system tools such as ending a call, webhook tools that call your server, and client tools run by the connected app. There are versions with rollback and built-in evaluations.
The launch post adds multilingual support (agents in 44 languages, with switching among English, Spanish, Hindi, French and Japanese mid-conversation available now and all 44 coming soon), knowledge bases that prefill context so the agent answers without pausing to search, and outbound campaign batches personalized per call.
The design argument: co-design versus cascade #
Cartesia argues that cascaded pipelines (speech-to-text, then an LLM, then text-to-speech) dominate enterprise work but can lose naturalness in turn-taking and interruptions without careful co-design, while pure speech-to-speech models "sound natural" but are "hard to control and less reliable for enterprise-grade tasks." Its answer is to tune its Ink speech-to-text, its Sonic text-to-speech and the harness as one system around an LLM you pick.
Two speech bubbles passing a bead between them to show turn-taking in a voice agent conversation
That framing matters to buyers because it is the central tradeoff in voice AI right now. Other recent launches pick different points on it; compare ElevenLabs' Eleven Realtime speech-to-speech models, Rime's Prism claim versus published latency, and Phonic-1's benchmark claims.
Reading the benchmarks #
Speech Agent Arena. Cartesia's chart lists Managed Agents first at 1115, followed by Google's Gemini 3.1 Flash Live Minimal at 1099, Gemini 3.8 Live at 1079, xAI's Grok Voice 3.0 at 1064, OpenAI's GPT-Live-1 variants near 1041 and 1037, and ElevenLabs entries at 1003 and 989. The Elo comes from blind human preference on tasks like booking appointments and ordering takeout. Elo gaps of 16 points to the next entry are real but small; ranks 2 through 6 are close.
EVA-Bench. ServiceNow's open benchmark covers customer service management, IT service management and HR scenarios, over 200 realistic cases. In Cartesia's chart, Managed Agents with GPT-5.5 scores 87.4 percent accuracy and 90.7 percent experience. The chart lists other systems such as Grok Voice 3.0 (73.7 percent accuracy, 71.2 percent experience), Gemini 3.8 Live (45.9 and 88.6) and GPT Realtime 2.1 (59.0 and 61.6). A ServiceNow director is quoted saying it "posted one of the strongest results we've seen to date."
Important caveats:
- Different reasoning models. Cartesia's footnote says benchmarks ran with GPT-5.5 as the reasoning model for Managed Agents, while some competitors are end-to-end speech-to-speech systems with their own reasoning. The comparison is of whole stacks, not voice layers.
- Vendor-presented chart. The EVA-Bench numbers are displayed on Cartesia's page. Check ServiceNow's published results before treating the table as final.
- Generic tasks. Appointment booking and takeout orders are not your calls. Domain vocabulary, noisy lines and regional accents can reorder rankings.
- Internal turn-taking study. The 85 versus 67 percent figure comes from Cartesia's own evaluations with callers; the method is not described.
For context on how Cartesia has ranked before, see our coverage of Sonic-3.6 on Artificial Analysis TTS boards and its separate research release on H-Net byte-level models, published the same week.
How does the Speech Agent Arena work? #
Artificial Analysis introduced the arena in an August 24, 2026 article. It evaluates speech-to-speech systems and cascaded systems as humans complete real-world tasks, measuring conversational preference and successful tool use. Because raters compare agents blind, the Elo reflects how people felt about the interaction, which rewards natural pacing and clear speech as much as raw correctness. That is a useful signal for customer-facing calls and a weaker one for back-office accuracy, which is why Cartesia pairs it with the accuracy-oriented EVA-Bench.
The two measures can disagree. In Cartesia's own EVA-Bench chart, Google's Gemini 3.8 Live scores a high experience number (88.6 percent) but low accuracy (45.9 percent), which shows a voice can feel natural while getting details wrong. Managed Agents is positioned as the system that scores high on both, the "Pareto frontier" language in the post. Whether that holds on your tasks depends on the LLM behind it and on your tools.
Why this launch matters #
Voice agents are moving from demos to phone lines that handle appointments, IT tickets and HR requests, and buyers are now asking vendors to prove quality with shared benchmarks instead of curated demos. Cartesia's launch is a useful case study because it publishes numbers on two different yardsticks, one that rewards how a conversation feels and one that rewards whether the task was done correctly. Even if you discount the vendor-presented charts, the structure of the claim is worth copying in your own evaluations: score preference and accuracy separately, disclose the reasoning model, and report the price per resolved call.
Questions buyers are asking #
Is this speech-to-speech? Not in the pure sense. It is a co-designed pipeline: Ink for listening, an LLM for reasoning and Sonic for speech, with the harness handling timing. Cartesia argues co-design recovers the naturalness that cascades lose.
Can I choose my own LLM? According to the docs, yes: "the LLM of your choice." The benchmark runs used GPT-5.5, so results with another model may differ, better or worse.
Is the price a gimmick? A 50 percent launch discount is a common way to win trials. Decide on the regular price and total cost per resolved call, and ask for a written price lock if you commit to volume.
What about reliability at scale? The post says Managed Agents has powered tens of millions of conversations. Treat this as a vendor claim and ask for references in your industry.
How does this compare with OpenAI's voice products? In Cartesia's chart, OpenAI's GPT-Live-1 variants score below the top entries on the Arena Elo, and GPT Realtime variants trail on EVA-Bench. These are Cartesia-presented comparisons; OpenAI's own positioning is covered in our post on GPT-Live-1.1.
What does it cost, really? #
The post says production voice agents run at $0.03 a minute at 50 percent off, so the regular price is implied to be about $0.06 a minute. Cartesia's page did not state, in the text we read, whether LLM tokens, telephony minutes and knowledge-base use are included. Since the product is designed around an LLM you choose, plan on a separate LLM line in your cost model, and confirm telephony charges. A useful metric is cost per resolved call, not per minute, because a slow or error-prone agent costs more minutes.
What to test before switching #
- Your own call set. Replay 100 real calls, including interruptions, accents, background noise and corrections.
- Tool-call reliability. Measure whether the agent calls the right API with the right arguments, and recovers from failures.
- Latency and turn-taking. Measure time to first audio and how it handles barge-in.
- Languages. Test each language you need now; some of the 44 are still coming soon.
- Compliance and data. Check call recording, retention, PII handling and regional hosting, particularly in finance and healthcare.
- Fallbacks. Decide how and when the agent hands off to a human.
Teams assembling their own stacks may also look at open building blocks such as Deepgram's acquisition of Daily and the Pipecat framework, or at AssemblyAI's TTS-1 for the speech-output layer.
What we could not verify #
We did not see ServiceNow's own published EVA-Bench table, so we cannot confirm Cartesia's reproduction of other systems' scores. Customer claims such as "tens of millions of conversations" were not independently confirmed, and we did not locate the matching Artificial Analysis leaderboard entry. We could not find the regular price or an itemized pricing page in the sources we read.
Related reading #
Figures are as published by Cartesia on October 29, 2026.