ElevenLabs has released two speech-to-speech models under the name Eleven Realtime, built for live voice conversations with AI. The headline specification is a 0.57-second response time, a threshold that makes machine dialogue feel close to human turn-taking. For companies running voice agents, that speed matters because pauses and overlaps decide whether callers stay engaged.
Two models for live voice dialogue
The lineup includes Eleven Realtime v1 Max and Eleven Realtime v1 Mini, positioned for different workloads. The vendor places v1 Max near the top left corner of the Artificial Analysis Speech to Speech leaderboard, a position that signals a combination of high accuracy with low latency. That placement is the main external reference point available for performance so far. It frames Max as the option for deployments where recognition quality cannot be sacrificed for speed.
The two models process speech directly into speech rather than chaining separate transcription, language, and synthesis stages. ElevenLabs describes v1 Mini as tuned for natural conversation at lower cost, while v1 Max is aimed at more complex tasks such as customer support. The distinction is functional: one variant prioritizes fluid everyday dialogue, the other handles structured service work with higher demands on understanding. Both are presented as real-time conversational models, not offline voice generators.
The timing reflects pressure on voice AI to move from demos to production telephony, support lines, and in-product assistants. Buyers now compare vendors on measured latency and accuracy rather than sample audio alone, and independent leaderboards such as Artificial Analysis have become a common reference. ElevenLabs is entering that comparison directly by citing leaderboard position alongside its own latency figure. The release therefore reads as an attempt to compete on operational metrics that procurement teams track.
What voice latency means for operations
For businesses using AI in sales and service, a sub-second voice loop changes how calls can be structured. A 0.57-second response allows an agent to confirm orders, qualify leads, book appointments, and answer routine questions without the long silences that prompt callers to hang up. Larger firms can place the Max version on support queues where mistakes are costly, while smaller teams can run Mini for inbound filtering and after-hours coverage. The practical difference is fewer abandoned conversations for the same traffic.
The choice between versions comes down to cost control and dialogue discipline. Pricing is set per minute plus token costs, with v1 Max priced above v1 Mini, and full rates are published separately. Mini includes the ability to distinguish a finished utterance from a backchannel sound such as "mm-hmm," which reduces false interruptions in natural speech. Buyers should test that endpointing behavior on their own call recordings, verify total per-minute plus token spend at expected volumes, and confirm that the OpenAI-compatible Realtime API and ElevenAgents platform fit existing telephony and CRM workflows. This release alone does not prove accuracy on accents, noise, or domain terminology.
The marker to watch is placement and retention on the Artificial Analysis Speech to Speech leaderboard over the next evaluation rounds. If Eleven Realtime v1 Max holds a top-left position while customers deploy it through ElevenAgents, the claim of accuracy plus low latency will carry weight in vendor selection. A slip in ranking or slow uptake would suggest the 0.57-second figure is not enough to win service workloads.
