ElevenLabs brings together conversation, reasoning, and speech synthesis with Eleven Realtime

ElevenLabs introduces Eleven Realtime, two conversational voice models capable of responding in under 620 ms, complete with reasoning and tool calling.

A voice agent can respond quickly without necessarily understanding a complex request. Conversely, models capable of elaborate reasoning often introduce a delay that disrupts the flow of conversation. With Eleven Realtime, introduced on October 20, 2026, ElevenLabs aims to bring these two qualities together in a new family of Speech-to-Speech models. Two versions are available: Mini, designed for natural exchanges at a lower cost, and Max, built for tasks that require reasoning and tool use. According to evaluations published by Artificial Analysis and cited by ElevenLabs, Max achieves 98.8% correct answers on spoken reasoning questions while starting to speak in an average of 0.57 seconds. A Voice Model Rather Than a Succession of Components

Designing an agent capable of holding a spoken conversation usually involves assembling several technologies. Speech must be recognized, the system must determine when the speaker has finished, a language model must prepare the response, and then a voice generator must render it. Each of these operations introduces its own delays and requires tuning to prevent the agent from cutting in or reacting too late.

ElevenLabs proposes to simplify this architecture with a family of models optimized end-to-end for conversation. Eleven Realtime directly combines audio understanding, request processing, and vocal response generation, with particular attention to conversational pacing. The company is therefore not simply presenting a new synthetic voice, but a system capable of conducting an interaction and, depending on the chosen model, triggering external tools.

Both variants support more than 90 languages and can use the more than 17,000 voices in the ElevenLabs catalog. This compatibility allows users to maintain a chosen vocal identity while adapting the model's capabilities to the application's constraints. However, the specific pricing details for the two new variants are not specified in the launch announcement. Mini Prioritizes Fluidity, Max Handles Complex Tasks

Eleven Realtime v1 Mini targets personal assistants, educational tutors, and interactive characters. Its optimization particularly concerns turn-taking management: recognizing that a person has truly finished their sentence, accepting an interruption, and continuing the exchange without systematically reacting to simple active listening cues like "mm-hmm." These behaviors, sometimes secondary in traditional language model benchmarks, become essential when the interaction relies solely on voice.

Eleven Realtime v1 Max is aimed more at use cases where the agent must understand a multi-step request, look up information, or perform an action. ElevenLabs specifically cites customer service and sales qualification. The agent can formulate tool calls to query external services or advance a procedure, rather than limiting itself to providing a spoken response.

This division corresponds to two different forms of efficiency. Mini prioritizes the natural flow of conversation and a reported lower cost, while Max seeks to preserve greater reasoning and execution capabilities. Both models can participate in voice exchanges, but their results diverge significantly as soon as tasks require decisions or interactions with applications. 98.8% Success Rate on Spoken Reasoning Questions

To compare the models, ElevenLabs relies on three public evaluations conducted by Artificial Analysis. The first, Big Bench Audio, presents 1,000 reasoning questions by voice and verifies the spoken answers. Max achieves 98.8% correct answers, compared to 75.9% for Mini. This difference of nearly 23 points illustrates the gap between the capabilities targeted by the two variants.

Latency measurements are particularly close. Under the same protocol, Max begins producing its vocal response after 0.57 seconds on average, compared to 0.62 seconds for Mini. According to ElevenLabs, no other model present in the evaluated ranking combines higher accuracy with a lower latency than Max. This position on the Pareto frontier describes a trade-off between two specific metrics, rather than superiority in all aspects of conversation.

The time to first sound does not correspond to the time required to complete a response, nor to the total time of an operation involving external services. The results measure a precise test environment; they therefore do not guarantee that every phone call or website interaction will receive a response in under 600 milliseconds. Interruptions and Tool Calls Change the Ranking

The second evaluation, Full-Duplex-Bench, examines conversation management rather than answer accuracy alone. It notably includes pauses, interruptions, turn-taking, and short vocal cues indicating that a person is listening without necessarily wanting to speak. Under this protocol, Mini reaches 95.8%, slightly ahead of Max at 94.1%. The model that performs best in reasoning is therefore not the one that scores highest on conversational mechanics.

The third evaluation, Speech Agent Arena, is based on live conversations compared blindly by human participants. Max achieves a score of 998, compared to 988 for Mini. These preference scores must be distinguished from success percentages: they reflect a ranking derived from comparisons, not a proportion of conversations deemed satisfactory.

The gap becomes clearly favorable to Max again when tools must be used. In the evaluated tasks, Max successfully completes 87.1% of requests involving tool calls, compared to 60.3% for Mini. This result highlights the commercial positioning of the two models: Mini may be sufficient for a large portion of routine exchanges, while Max appears better suited for procedures requiring multiple operations. However, the success measured in these tests does not guarantee that an agent will correctly execute a real-world workflow without supervision. Two Access Points: ElevenAgents and an OpenAI-Compatible API

Both models are available in ElevenAgents, ElevenLabs' platform for