ElevenLabs has launched two new speech-to-speech AI models, Eleven Realtime v1 Mini and Eleven Realtime v1 Max, designed to make AI voice agents faster, more natural and easier for developers to build. Announced on October 20, 2026, the models combine speech recognition, conversational reasoning, turn-taking and voice generation into a single managed system.
The launch targets businesses and developers building AI assistants, customer-support agents, virtual tutors and other applications that need to communicate with people in real time. Rather than requiring developers to configure separate speech recognition, language model and text-to-speech components, Eleven Realtime brings these capabilities together in one model choice.
According to ElevenLabs, both models support more than 90 languages and can use voices from its library of more than 17,000 options. The company also says its new Realtime API is compatible with OpenAI’s Realtime API, allowing developers with existing integrations to test the models by changing their endpoint and model configuration.
The two models serve different purposes. Eleven Realtime v1 Mini is designed for affordable, natural conversations, while v1 Max targets more complex tasks that require reasoning and multiple tool calls.
Key takeaways
- Two new models: ElevenLabs introduced Eleven Realtime v1 Mini and v1 Max.
- Speech-to-speech architecture: The models integrate the major components required for a conversational voice agent.
- Fast responses: ElevenLabs reports average time-to-first-audio figures of 0.62 seconds for Mini and 0.57 seconds for Max.
- Different use cases: Mini targets assistants, tutors and interactive characters, while Max is designed for customer support, lead qualification and multi-step workflows.
- 90+ languages: Both models support more than 90 languages.
- 17,000+ voices: Developers can choose from ElevenLabs’ voice library.
- Developer compatibility: The Realtime API follows OpenAI’s Realtime API specifications over WebSocket.
- Independent evaluation: Artificial Analysis assessed the models across speech reasoning, conversational dynamics and human preference benchmarks.
What is ElevenLabs Eleven Realtime?
Eleven Realtime is a new category of speech-to-speech models from ElevenLabs. It is designed to simplify the process of building AI agents that can listen, understand spoken requests and respond verbally with minimal delay.
Traditional voice-agent systems often rely on several connected components. A speech recognition model converts incoming audio into text, a language model generates a response, and a text-to-speech model converts the response back into spoken audio. Separate systems may also manage pauses, interruptions and the detection of when a person has finished speaking.
Although this modular architecture offers flexibility, developers must configure and coordinate the individual components. Differences between them can affect response time, conversational flow and the overall user experience.
Eleven Realtime aims to reduce this complexity by combining the core voice-agent components into a single managed model configuration.
For a developer, that means selecting a Realtime model instead of separately choosing every component of the voice stack. The agent can still use its configured instructions, voice, tools and knowledge base, while ElevenLabs manages the underlying real-time conversational components.
The goal is to make it easier to build voice agents that can respond naturally, recognise when someone has finished speaking and handle interruptions without requiring extensive manual tuning.
Eleven Realtime v1 Mini vs v1 Max
The two models are designed for different priorities rather than representing identical options at different sizes.
<box gap={3}> <row align=start gap={3}> <AsyncImage query=”minimal modern AI voice assistant sound wave microphone illustration” aspectRatio=”3:4″ width=”104px” /> <box flex=”1″ gap={1}> <title size=”md”>Eleven Realtime v1 Mini</title> <badge color=”success”>Lower cost</badge> Mini prioritises fast, natural conversation at a lower cost. It is aimed at AI assistants, interactive characters, tutors and applications where conversational flow matters more than advanced multi-step reasoning.
<text color="secondary" size="sm">Best suited to everyday conversational experiences.</text>
</box>
</row> <divider color=”subtle” /> <row align=start gap={3}> <AsyncImage query=”professional customer support voice AI headset sound wave abstract technology illustration” aspectRatio=”3:4″ width=”104px” /> <box flex=”1″ gap={1}> <title size=”md”>Eleven Realtime v1 Max</title> <badge color=”info”>Complex tasks</badge> Max is designed for more demanding conversations involving detailed instructions, tool calls and multi-step workflows. It targets customer support, lead qualification and agents that need to look up information or take actions during a conversation.
<text color="secondary" size="sm">Best suited to task-oriented voice agents.</text>
</box>
</row> </box>
The practical difference is that Mini is intended to deliver a cost-conscious conversational experience, whereas Max is designed for situations in which the agent must reason through more complicated requests and use tools reliably.
Businesses should evaluate both models against their own workloads. A simple appointment assistant may not need Max’s additional reasoning capabilities, while an agent that checks several systems before resolving a customer issue may benefit from it.
How fast are the new Eleven Realtime models?
Response speed is important for voice agents because even short delays can make a conversation feel unnatural. A person expects a spoken response to follow a question without the extended pause often associated with traditional automated systems.
ElevenLabs published the following average time-to-first-audio results from Artificial Analysis’ evaluation.
| Metric | Eleven Realtime v1 Mini | Eleven Realtime v1 Max |
|---|---|---|
| Average time to first audio | 0.62 seconds | 0.57 seconds |
| Big Bench Audio speech-reasoning accuracy | 75.9% | 98.8% |
| Full-Duplex-Bench score | 95.8% | 94.1% |
| Speech Agent Arena score | 988 | 998 |
| Tool-use task completion | 60.3% | 87.1% |
Source: ElevenLabs launch announcement. Figures are reported by ElevenLabs from Artificial Analysis evaluations; benchmark scores use different measurement scales and should not be compared directly across tests.
The reported time-to-first-audio measures how quickly a model begins producing speech. It does not mean the model completes every answer within that time. Longer questions, tool calls and complex reasoning can take additional time.
The results also suggest that the fastest model is not necessarily the best option for every task. Mini performed slightly better on the reported Full-Duplex-Bench score, which evaluates conversational dynamics, while Max scored higher on speech reasoning and tool-use completion.
For developers, the choice depends on whether the application primarily needs natural conversation, complex decision-making or a balance between the two.
What do the benchmark results mean?
ElevenLabs says Artificial Analysis independently evaluated the models across three public benchmarks. Each examines a different part of the voice-agent experience.
1. Big Bench Audio: Can the model reason through spoken questions?
Big Bench Audio evaluates a model’s ability to answer reasoning questions presented through audio.
In the reported test of 1,000 spoken reasoning questions, Eleven Realtime v1 Max achieved 98.8% accuracy, while v1 Mini achieved 75.9%.
This difference is particularly relevant for agents that must interpret a spoken request correctly before taking action. Customer-support systems, for example, may need to understand a complicated question, identify the relevant information and provide a suitable answer.
However, benchmark accuracy is not the same as guaranteed real-world reliability. Results can vary with the type of question, the language used, the quality of the audio and the complexity of the actual workflow.
2. Full-Duplex-Bench: Can the agent manage natural conversation?
Full-Duplex-Bench evaluates conversational behaviour, including pauses, interruptions, turn-taking and short verbal acknowledgements.
Eleven Realtime v1 Mini scored 95.8%, while v1 Max scored 94.1% in the reported evaluation.
This matters because spoken conversation is not simply a sequence of questions followed by complete answers. People interrupt, change their minds, pause to think and use brief acknowledgements such as “mm-hmm” while another person is speaking.
A voice agent that mistakes an acknowledgement for the end of a turn may interrupt unnecessarily. One that waits too long may make a conversation feel slow.
These benchmark results are intended to assess how well models manage such conversational details.
3. Speech Agent Arena: Which model do people prefer?
The Speech Agent Arena uses blind comparisons of live conversations to assess human preference and task success.
ElevenLabs reported scores of 998 for v1 Max and 988 for v1 Mini.
Human-preference evaluations can capture aspects of voice interaction that are difficult to represent with a single accuracy metric, including how well an agent responds to a request and how natural the exchange feels.
Nevertheless, benchmark results do not establish that every user will prefer Max or that it will perform better in every application. Developers should test their intended use cases before choosing a production model.
Why speech-to-speech AI matters
Voice agents are becoming an important interface for customer service, education, accessibility and software applications. Yet traditional voice systems can feel mechanical when they respond slowly, misunderstand interruptions or require callers to follow rigid scripts.
Speech-to-speech systems aim to make these interactions more fluid.
Instead of treating spoken input as a transcription task followed by a separate text conversation and speech synthesis step, a speech-to-speech system is designed around the complete audio interaction.
This can help developers create experiences where users speak naturally, ask follow-up questions and interrupt when necessary.
Potential applications include:
- Customer support: Answering questions, looking up account information and helping customers resolve common issues.
- Sales and lead qualification: Asking questions, gathering requirements and determining whether a lead meets defined criteria.
- Education: Supporting spoken tutoring, language practice and interactive learning.
- Virtual assistants: Helping users complete everyday tasks through natural conversation.
- Interactive characters: Powering spoken interactions in games, entertainment and virtual experiences.
- Accessibility: Offering voice-based interfaces for people who find typing or conventional screen-based interaction difficult.
These are potential uses rather than guaranteed outcomes. The reliability of any deployed system will depend on its design, access to relevant information, safeguards and ability to handle exceptional cases.
How Eleven Realtime simplifies voice-agent development
Before a real-time voice model can be deployed, developers typically need to make decisions about several components.
| Component | Traditional development challenge | Eleven Realtime’s approach |
|---|---|---|
| Speech recognition | Select and configure an audio transcription system | Managed as part of the Realtime model |
| Language model | Choose a model suited to the application’s tasks | Included in the managed system |
| Text-to-speech | Configure voice generation and output | Integrated into the model |
| Turn-taking | Tune when the agent should speak or listen | Optimised as part of the Realtime experience |
| Interruptions | Handle users speaking while the agent responds | Designed to support natural interruption handling |
| Voice selection | Connect a suitable voice to the application | Access to the ElevenLabs voice library |
| Tool use | Build workflows for external lookups and actions | Particularly supported by the Max model |
This consolidated approach could reduce integration work and the number of separate decisions developers need to make before testing a voice agent.
It does not eliminate the need for engineering. Developers still need to define agent instructions, configure tools, manage permissions, test failure cases and ensure that the agent behaves appropriately when it encounters ambiguous requests or unavailable information.
The key difference is that the underlying real-time voice stack is more tightly managed by the provider.
Eleven Realtime API supports OpenAI-compatible integrations
ElevenLabs is also targeting developers who already have applications built around OpenAI’s Realtime API.
The company says its Realtime API follows the OpenAI Realtime API specifications and uses WebSocket connections. This allows developers to test Eleven Realtime by changing the endpoint and model name in an existing compatible integration, rather than rebuilding the entire application from scratch.
The potential benefit is a shorter evaluation process for teams that want to compare voice models without making major changes to their infrastructure.
However, API compatibility should not be interpreted as a guarantee that every feature, parameter or workflow will behave identically. Developers should review the official ElevenLabs documentation, check the supported capabilities and test their applications before switching production traffic.
ElevenLabs also offers the models through ElevenAgents, its platform for building and operating voice agents. The platform includes tools for deployment across supported channels, knowledge bases, workflows, guardrails, testing and monitoring.
This gives businesses a managed route to deployment, while the API provides a path for teams building their own applications.
Language support and voice selection
ElevenLabs says Eleven Realtime supports more than 90 languages and can use any of the more than 17,000 voices in its voice library.
For businesses operating internationally, multilingual support could make it easier to build voice experiences for customers who prefer languages other than English.
A customer-service provider, for example, could develop an agent that supports different language groups. An educational application could use spoken interactions for language learning, while a global business could explore voice automation across different markets.
Voice selection also matters. The sound and identity of a voice influence whether an interaction feels appropriate for its intended audience. A professional support agent may require a different voice from a children’s learning assistant or an interactive game character.
Still, broad language support does not guarantee equal accuracy or naturalness in every language and accent. Businesses should test the model with representative speakers, real customer questions and relevant terminology before deploying it at scale.
What does the launch mean for customer-support businesses?
Customer support is one of the clearest target markets for Eleven Realtime v1 Max.
A typical support conversation may involve more than answering a simple question. The agent might need to understand the customer’s problem, search a knowledge base, retrieve an order, apply a policy and then explain the result.
That kind of multi-step process places different demands on a voice system than a casual conversation.
ElevenLabs reported that v1 Max completed 87.1% of tool-use tasks in the cited evaluation, compared with 60.3% for v1 Mini. The company recommends Max for workloads that rely heavily on tool calling.
For businesses, stronger tool-use performance could be relevant when an agent must retrieve information or take actions during a call. It is still important to design clear limits around those actions, especially when they affect accounts, payments, personal data or customer entitlements.
A voice model should not be given unrestricted authority simply because it performs well on a benchmark. Production systems need authentication, permission checks, escalation paths and monitoring.
Pricing and availability
ElevenLabs announced Eleven Realtime on October 20, 2026, with access through ElevenAgents and its Realtime API. The company says the models are rolling out to users, although availability may vary by workspace.
The launch announcement does not provide a single fixed price for all use cases. Developers should check the ElevenLabs pricing information and current API documentation for applicable charges and usage conditions.
The two models are positioned differently on cost: Mini is intended to offer a lower-cost conversational option, while Max is designed for more demanding tasks and costs more.
When estimating the cost of a deployed agent, businesses should consider more than the model rate. Total operating costs can also depend on conversation duration, usage volume, tool calls, integrations, monitoring and the engineering required to maintain the system.
How Eleven Realtime compares with traditional voice-agent architectures
Traditional voice agents often combine an automatic speech recognition system, a language model and a text-to-speech engine. Developers can choose each component independently, which offers flexibility but also creates integration work.
Eleven Realtime takes a more integrated approach by providing a managed model configuration for the full conversational experience.
| Area | Modular voice-agent stack | Eleven Realtime |
|---|---|---|
| Component selection | Separate systems must be selected | Managed Realtime model choice |
| Configuration | Developers tune several components | Fewer underlying components to configure |
| Customisation | Potentially extensive at component level | More limited control over managed components |
| Conversational behaviour | Depends on integration and tuning | Models are optimised for real-time conversation |
| Complex tool use | Depends on the chosen architecture | Max is specifically designed for demanding workflows |
| Existing API migration | May require integration work | OpenAI-compatible Realtime API offers a migration path |
Neither architecture is universally better. Teams that need highly customised components or a specific language model may prefer a modular stack. Teams prioritising faster implementation and a managed conversational experience may find Eleven Realtime easier to evaluate.
The right choice depends on technical requirements, cost, control and measured performance in the intended application.
What developers should consider before adopting Eleven Realtime
The launch offers a simpler route to building voice agents, but companies should evaluate several factors before committing to a model.
Test real conversations. Use representative customer questions, different speaking styles, interruptions and noisy audio conditions rather than relying only on public benchmarks.
Compare Mini and Max. Choose the least expensive model that meets the application’s requirements, and evaluate whether Max’s reasoning and tool-use advantages justify the additional cost.
Check language performance. Test the languages and accents your users actually speak.
Secure connected tools. Restrict the agent to the actions and data it needs, and require authentication for sensitive operations.
Plan for escalation. The agent should be able to acknowledge uncertainty and transfer complicated or high-risk cases to a human.
Monitor ongoing performance. Track latency, task completion, user satisfaction, errors and the frequency of human handoffs after deployment.
These steps are essential because an agent that performs well in a controlled benchmark may behave differently in a production environment with real customers, changing data and unexpected requests.
Frequently asked questions
1. What are ElevenLabs Eleven Realtime models?
Eleven Realtime is a family of speech-to-speech AI models designed to power voice agents. They combine speech recognition, conversational reasoning, turn-taking and speech generation into a managed system.
2. What are the two new Eleven Realtime models?
The two models are Eleven Realtime v1 Mini and Eleven Realtime v1 Max. Mini targets fast, natural and lower-cost conversations, while Max is designed for complex tasks and multi-step tool use.
3. How fast are Eleven Realtime models?
ElevenLabs reports average time-to-first-audio figures of 0.62 seconds for Mini and 0.57 seconds for Max in its published evaluation. These figures measure when speech begins, not the time required to complete every answer.
4. Which Eleven Realtime model is better for customer support?
Eleven Realtime v1 Max is the company’s recommended option for complex customer-support workflows, lead qualification and applications that need multiple tool calls. Mini may be sufficient for simpler, lower-cost interactions.
5. How many languages do Eleven Realtime models support?
ElevenLabs says both models support more than 90 languages. Businesses should test the actual languages, accents and terminology used by their customers before deployment.
6. Can developers use Eleven Realtime with existing OpenAI Realtime API applications?
ElevenLabs says its Realtime API is compatible with OpenAI’s Realtime API specifications over WebSocket. Developers may be able to test the models by changing their endpoint and model configuration, but they should validate feature compatibility before migrating production systems.
7. Are Eleven Realtime models available through an API?
Yes. ElevenLabs offers access through ElevenAgents and its Realtime API. Availability may vary during the rollout, so developers should consult the official documentation.
8. How much do Eleven Realtime models cost?
Pricing depends on the applicable model and usage terms. Mini is positioned as the lower-cost option, while Max costs more. Developers should check current pricing directly with ElevenLabs before estimating deployment costs.
9. What makes speech-to-speech AI different from traditional voice assistants?
Speech-to-speech systems are designed around the complete audio conversation rather than treating transcription, text-based reasoning and voice generation as entirely separate stages. The goal is to improve conversational flow, responsiveness and interruption handling.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
