One model where three services used to sit
A conventional voice agent is an assembly: one service transcribes the caller, a language model composes the reply, a second voice service reads it out. Each joint in that chain adds delay and a place for tone and timing to go wrong. Eleven Realtime removes the joints. In ElevenLabs' product description, understanding, reasoning, turn-taking and voice are tuned together as one end-to-end system, so the model that hears an interruption is the same one that decides to stop talking.
The detail that shows what this design is meant to buy is a small one. Mini is presented as able to tell the difference between a caller who has finished a sentence and one who is only saying a backchannel such as mm-hmm to show they are still listening. In a stitched pipeline, that judgment belongs to a separate voice-activity component with its own thresholds. Inside one model it is part of the conversation itself. That is a capability claim, not a demonstrated reliability rate across real-world call conditions. Whether it survives a noisy call centre line, a strong accent and a caller talking over hold music is a different question, and launch demos do not answer it.
Read the two numbers with their source attached
The launch leads with two numbers for Max, and they need to be kept in three separate layers. First, ElevenLabs reports 98.8% on Big Bench Audio and a response time of about 0.57 seconds, and it reports a position on the Artificial Analysis speech-to-speech leaderboard under which no listed competitor is both faster and more accurate. That leaderboard position is a Pareto claim about the listed set, not a blanket claim about speed or accuracy in every setting. Second, ElevenLabs says, via AI Parabellum, that Artificial Analysis ran the evaluations independently. That is a claim about how the evaluations were administered. Third, no separately published independent reproduction with fully documented comparable settings exists in the sources reviewed for this article.
Big Bench Audio is an audio reasoning test under benchmark conditions. It does not establish performance on noisy lines, across accents, through interruptions and tool-call failures, or over the length of a real support call. The latency figure likewise needs its measurement conditions before any cross-provider comparison is defensible. A chart that travels well on social media says nothing about how the model handles your product names, your refund policy, or the third minute of an angry call. Those are the tests that decide whether a support line gets better.
Two meters instead of one flat rate
Mistral also launched Voxtral 3 on 20 October, and its published voice-agent pricing states the condition plainly: $0.04 per minute with Mistral Small 4 as the reasoning model, and $0.08 per minute with any other reasoning model. We covered that launch in [our Voxtral 3 piece](https://alfw.ai/article/mistral-voxtral-3-voice-agent-stack/).
ElevenLabs' structure, per the reported documentation, runs on two meters. Through the Realtime API, the charge combines minutes of connected audio with token usage. Inside ElevenAgents, the structure is a standard per-minute rate plus LLM token costs. Max carries a higher token rate than Mini. Exact rates were not independently verified for this article, and this article does not claim they are absent from ElevenLabs' site.
Two-meter billing can be estimated from expected call duration and expected token use. What it does not give a buyer, on its own, is a defensible dollar-for-dollar comparison with a flat per-minute rate. That comparison requires the exact rates and the buyer's own workload — how long calls run, how verbosely the agent reasons, and how often tools are called — and neither provider should be assumed cheaper before that calculation is done.
The compatibility is the distribution strategy
ElevenLabs presents the Realtime API as compatible with OpenAI's format, potentially lowering the work involved in testing it. Per AI Parabellum's account of the documentation, clients can reuse familiar request and event structures with endpoint and model changes. That is a distribution advantage, not a guarantee of drop-in compatibility. Teams still need to verify authentication, supported events, tool calls, audio transport and interruption handling against their own integrations before treating a switch as a configuration change.
That distinction is the commercially important one in this launch. An API format a team already knows lowers the cost of running a side-by-side test on recorded calls. It does not remove the test.
Limits, below the headline
The limits that follow are carried as AI Parabellum's account of ElevenLabs' documentation, as reported on 20 October 2026. The Realtime API connects over WebSocket only. Custom mode is retained for setups that use a custom language model, Amazon Bedrock, or workflow nodes that override managed settings. HIPAA support and Zero Retention Mode are Max-only; the availability of HIPAA support does not by itself make an application HIPAA-compliant. Max carries a minimum token limit of 1,024. Max is unavailable in the India and Singapore data-residency regions. Rollout is gradual, so launch-day availability and a given workspace may disagree for a while.
Languages need the same care. Coverage in more than 90 languages is ElevenLabs-reported coverage, not evidence of equal quality across those languages. Multilingual buyers should test their own languages and accents individually, on their own recordings.
A note on sources
ElevenLabs' announcement, documentation and pricing page were inaccessible to our verifier. Product details and restrictions in this article are attributed to The Decoder, AI Parabellum and Oossa. Exact current pricing and documentation terms were not independently confirmed against ElevenLabs' website.
Sources
- [1] The Decoder, 'Elevenlabs releases two new speech-to-speech models designed for real-time AI conversations', 20 October 2026Read source
- [2] AI Parabellum, 'ElevenLabs Eleven Realtime: New Voice Agent Models', 20 October 2026 — a launch write-up citing ElevenLabs' announcement and Realtime documentation as its primary sourcesRead source
- [3] Oossa, 'Elevenlabs launches two real-time speech-to-speech models', 20 October 2026Read source
- [4] Mistral AI, 'Voxtral 3' announcement and pricing table, 20 October 2026 (comparison figure only: $0.04/min with Mistral Small 4, $0.08/min otherwise)Read source