FULL STORY

Cartesia Managed Agents: From Arena Leader to Enterprise Launch

After topping Artificial Analysis's new Speech Agent Arena blind-test leaderboard with an Elo of 1115, Cartesia officially launched Managed Agents on October 30 as an end-to-end enterprise voice agent platform.

2026-10-28 ~ 2026-10-30 · 2 episodes · 12 posts

Episode 1 · Cartesia Managed Agents tops Speech Agent Arena with Elo 1115 (2026-10-28, 7 posts)

Artificial Analysis has launched Speech Agent Arena, a blind-preference leaderboard for voice agents, where Cartesia Managed Agents ranks first with an Elo of 1115, ahead of Gemini Flash Live and Cartesia's own older product. The product also leads on tool-calling task success rate, and is priced at $0.06 per minute.

Confirmed

  • The leaderboard is based on human blind testing: users have real-time voice conversations with two hidden systems and pick the better one, with task scenarios including real voice calls like booking a dentist appointment or ordering takeout
  • Cartesia Managed Agents has an Elo of 1115 (95% CI 1086–1144, 685 samples), 98 points above its own older product, which sits at Elo 1017
  • For tasks requiring tool calls, "task success rate" measures the share of conversations ending with a correct final tool call; Cartesia Managed Agents achieves 91.7%
  • Artificial Analysis plots Elo and task success rate on the same scatter chart, showing Cartesia Managed Agents improving on both axes simultaneously
  • Pricing: $0.06 per minute of call time (equivalent to $3.60/hour), with LLM usage billed separately by model

Why it matters

  • Voice agents are becoming a new focal point of competition in voice AI, and this leaderboard is the first to systematically compare real-time conversational voice capabilities via human blind-preference testing
  • Cartesia leads on both Elo preference and tool-calling success rate, indicating strengths in both conversational experience and task completion rather than excelling on a single dimension

Episode 2 · Cartesia Launches Managed Agents, an End-to-End Enterprise Voice Agent Platform (2026-10-30, 5 posts)

On October 30, voice model company Cartesia released Managed Agents, an end-to-end voice agent platform for enterprises. The company says the platform has topped both Artificial Analysis's Speech Agent Arena and ServiceNow's EVA-Bench.

Confirmed

  • The platform takes a fully in-house, full-stack approach: proprietary speech recognition (Ink 2) and speech synthesis (Sonic 3.6), with the platform taking over the entire conversation loop—including speech recognition, turn-taking detection, interruption handling, and telephony; users bring their own LLM.
  • The core design philosophy is to co-design the voice model, inference, harness, context, and deployment tooling as one integrated system rather than assembling components from different vendors, jointly optimizing the voice model with conversation orchestration (the harness).
  • Turn detection is built in, eliminating the need for a separate turn-taking model; per @AiBreakfast, developers only need to supply a prompt to build phone-in, phone-out applications.

Why it matters

  • The release represents a shift in the voice agent stack from "assembling best-of-breed components" to "co-designing model and orchestration layer," which could reshape how enterprise voice agents are built and how the competitive landscape evolves.
  • The twin benchmark wins (per the company) are Cartesia's main argument that a fully integrated, in-house stack beats component assembly.