In-depth analysis · About 4 min read
Eleven Realtime: choosing a voice agent requires more than the fastest first sound
Mini and Max combine a voice-agent interface, but reasoning, interruptions and tool completion remain different tests.
- Published here
- Source published
- Source checked
ElevenLabs announced Eleven Realtime v1 Mini and Max on October 20. Its full announcement describes speech-to-speech models available through ElevenAgents or a separate WebSocket API. The practical question is which conversational tasks can be handed over, rather than whether one latency number makes a complete service reliable.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- The announcement reports 98.8% versus 75.9% on spoken reasoning and 87.1% versus 60.3% on tool tasks, for Max and Mini respectively. These are reported evaluation results, not AISubHub tests.
- The API uses WebSocket, not WebRTC. ElevenAgents provides platform features separately; compatibility with an existing SDK does not automatically migrate an application.
What changes when the voice stack is packaged together
A conventional voice application connects speech recognition, a language model and speech generation, with additional decisions about pauses and interruptions. ElevenLabs presents its models as a more integrated conversational interface. That can reduce the number of components an application team coordinates, but the announcement does not establish that every internal stage has disappeared. A developer still defines the business rules and what the agent may do. For example, answering a question about an appointment and actually changing that appointment are different responsibilities: a convincing spoken answer is not evidence that the booking system accepted an authorized update.
Four measurements answer four different questions
The announcement attributes its evaluation to Artificial Analysis. Big Bench Audio asks 1000 spoken reasoning questions: Max is reported at 98.8% and Mini at 75.9%. Tool-task success is separately reported as 87.1% and 60.3%. Yet Mini’s Full-Duplex-Bench score, 95.8%, exceeds Max’s 94.1%; that benchmark concerns conversational dynamics such as turn-taking. The Speech Agent Arena values are 998 and 988, not percentages. Our reading is that these results support task-specific comparisons, not a universal winner. We did not independently audit evaluation samples, uncertainty or the conditions behind each reported number.
First audio is the beginning of the response
The reported average time until speech begins is 0.57 seconds for Max and 0.62 seconds for Mini. This is not the time to finish an answer, complete a tool call or receive the result over every telephone network. Consider a hypothetical caller who corrects an address halfway through a sentence. An early reply can feel responsive while still committing the old address. A useful application test therefore measures interruption recovery and whether the final external state matches the correction, alongside first-audio time. This example is an evaluation proposal, not a claim that either model made that error.
An API-compatible connection is only part of migration
ElevenLabs describes an OpenAI Realtime-compatible API and shows an existing SDK connecting after changing the endpoint and model. It also explicitly says the connection is WebSocket-only: an application using WebRTC must change that transport. The standalone API runs separately from ElevenAgents, whose channels, knowledge base, workflows, guardrails and monitoring belong to the platform path. Consequently, a team should inventory the features it currently relies on before assuming that connection compatibility means feature parity. Authentication, application-side tool authorization, error handling and recovery still need verification in the actual integration; the sample code alone is not a production acceptance test.
Choose a task boundary before choosing a model
The vendor positions Mini for routine conversation and Max for more demanding, tool-heavy workflows, while stating support for more than 90 languages and a library of over 17000 voices. A sensible pilot keeps the task, language, audio conditions and permissions fixed when comparing them. Record correct answers, completed authorized actions, interruption handling and total task time; do not turn the Arena number into a success percentage. Pricing and access must be checked separately because the announcement does not establish a free or open-source entitlement. A larger voice library creates more presentation choices, not proof of better reasoning, consistent dialect coverage or permission to use every voice in any commercial context.
Why it matters
AISubHub’s analysis is that model selection should separate conversational comfort, correct decisions and successful authorized actions. The best trial uses the same tasks and records all three, including what happens after an interruption.
Limits and availability
We read the complete announcement, including integration instructions. We did not independently inspect the underlying benchmark runs, execute the API, buy credits or test voices. The article supplies no verified unit price or general free entitlement; language and voice counts do not guarantee equal performance or unrestricted voice rights.
Sources and evidence
Source published: · Source checked:
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.