ElevenLabs bets on one-model voice agents, with 98.8% on a speech reasoning test

ElevenLabs has launched Eleven Realtime, a family of speech-to-speech models in which a single model acts as the whole voice agent. The top model, v1 Max, answered 98.8% of spoken reasoning questions correctly in an Artificial Analysis test, according to the company.

Two models, one API that speaks OpenAI’s dialect

Until now, a natural-sounding voice agent meant assembling separate parts: speech recognition, voice activity detection, turn-taking, a language model and text to speech. ElevenLabs says that stack is a project in itself and needs a lot of testing and tuning. Eleven Realtime is meant to remove that work. Each model is a complete agent, and both support more than 90 languages and any of the 17,000+ voices in the ElevenLabs library.

There are two versions:

• Eleven Realtime v1 Mini targets natural conversation at low cost. ElevenLabs recommends it for AI assistants, teachers and interactive characters and says it “will be sufficient for most use cases.”
• Eleven Realtime v1 Max is aimed at customer support, lead qualification and multi-step workflows where the agent must follow a process and reason through steps.

There are two ways in. The company recommends ElevenAgents, its platform: create or open an agent and pick a Realtime model from a dropdown. That brings phone, web, WhatsApp and contact-center channels, workflows, a knowledge base, guardrails, testing and monitoring.

The second route is a Realtime API that follows the OpenAI Realtime specification over WebSocket. Developers with existing OpenAI Realtime code can, per ElevenLabs, change the endpoint and model name. The example in the post points OpenAI’s agents SDK at wss://api.elevenlabs.io/v1/convai/realtime with the model eleven_realtime_v1_mini.

There are limits. The API supports WebSocket only, so WebRTC integrations must move to WebSocket first. It also runs separately from ElevenAgents, so teams that need channels or a knowledge base should build on ElevenAgents from the start. A migration guide covers differences from OpenAI’s API.

Benchmark results, and what they leave out

ElevenLabs cites Artificial Analysis, which it says evaluated both models independently on three public benchmarks.

• Speech reasoning (Big Bench Audio, 1,000 spoken questions): Max 98.8%, Mini 75.9%.
• Time to first audio: Max 0.57 seconds on average, Mini 0.62 seconds.
• Conversational dynamics (Full-Duplex-Bench): Mini 95.8%, Max 94.1%.
• Speech Agent Arena (blind live comparisons): Max 998, Mini 988.
• Tool-use tasks: Max 87.1%, Mini 60.3%.

The gap in that last pair is the practical takeaway. ElevenLabs itself recommends Max for heavy tool calling, which makes Mini’s lower price a real trade-off for agents that must act, not just talk.

The announcement does not give pricing for either model, and it does not name the competing models in its charts in the text we could read. It also does not describe the architecture, context limits or rate limits. The figures are ElevenLabs’ presentation of Artificial Analysis results; we have not seen the underlying leaderboard entries.

Numbers in context: A crowded speech-to-speech field

Artificial Analysis launched a Speech to Speech Index in June 2026 that weights three datasets equally: Big Bench Audio, a Full Duplex Bench subset and the agentic τ-Voice test. In its announcement, Artificial Analysis said Grok Voice Think Fast 1.0 led Big Bench Audio at 97.1%, with top scores tightly clustered. A 98.8% result for Max would sit above that figure, although the two numbers come from different dates and we could not confirm Max on the live leaderboard.

On speed, the same Artificial Analysis post listed time to first audio of 0.82 seconds for GPT-Realtime-1.5, 1.25 seconds for Grok Voice Think Fast 1.0 and 2.33 seconds for GPT-Realtime-2 (High). Against those, 0.57 and 0.62 seconds are faster, and ElevenLabs argues that Max avoids the usual trade-off in which the most accurate models answer slowest.

One difference matters to buyers. The Artificial Analysis index scores agentic performance with τ-Voice, where it reported Grok Voice Think Fast 1.0 leading at 52.1% and all models below 53%. ElevenLabs reports tool use through the Speech Agent Arena instead, so its 87.1% is not directly comparable with those τ-Voice numbers.

A natural next step for the voice specialist

ElevenLabs built its name on synthetic voices and has been pushing into agents, as in our ElevenLabs review. Eleven Realtime moves the company from supplying one layer of the voice stack, the voice, to supplying the whole conversation. Making the API compatible with OpenAI’s Realtime specification lowers the cost of switching for anyone already building there. The company also keeps drawing developers into its community through events such as the Chroma Awards, where it added $22,000 in prizes.

For readers, the immediate effect is a simpler path to a voice agent: pick a model, pick a voice, and skip the tuning of separate components. What is still missing is cost, which will decide whether Mini really is the default for most deployments, and independent testing of both models outside ElevenLabs’ own chosen benchmarks.