ElevenLabs Eleven Realtime Review: Pricing, Latency Limits and Mistral Voxtral 3 Comparison

ElevenLabs released Eleven Realtime on October 20, 2026, with v1 Mini and v1 Max models built for end-to-end speech-to-speech voice agents. Both versions target developers building assistants, support agents, and interactive characters that respond in real time without separate STT and TTS steps.

The service runs over WebSocket and reports average time-to-first-audio of 0.57 seconds for Max and 0.62 seconds for Mini. Tool-use success reaches 87.1 percent on Max and 60.3 percent on Mini according to the launch benchmarks. These numbers set the baseline for production decisions on cost and capability.

Eleven Realtime Pricing and Limits

Eleven Realtime charges $0.08 per connected audio minute on standard plans, with burst traffic billed at $0.16 per minute. Token costs differ sharply by model. Mini runs at roughly $0.20 per million input tokens and $6 per million output tokens. Max costs $1.80 per million input and $9 per million output. Free accounts include 12 to 14 minutes of audio before overage fees apply, and telephony carrier charges remain separate on all tiers.

Usage limits include a minimum 1,024-token context for Max and no fixed per-minute agent rate beyond the audio-plus-tokens structure. Max supports Zero Retention and HIPAA with US or EU data residency; Mini does not. Max is unavailable for India or Singapore residency options. Both models support 90-plus languages and more than 17,000 voices, but production teams must monitor token spend when conversations lengthen or require frequent tool calls.

Typical costs for a support agent handling 500 minutes per month start near $40 in audio fees plus token usage that can add another $30 to $120 depending on model and output length. Teams running high-volume qualification flows often test Mini first, then switch select calls to Max for complex reasoning.

Latency Benchmarks and Mistral Voxtral 3 Comparison

Eleven Realtime Max scores 98.8 percent on Big Bench Audio and 94.1 percent on Full-Duplex-Bench, while Mini reaches 75.9 percent and 95.8 percent respectively. Human preference scores in the Speech Agent Arena sit at 998 for Max and 988 for Mini. These results come from ElevenLabs-reported tests and an independent Artificial Analysis evaluation at launch.

Mistral Voxtral 3, released the same day, offers a LiveKit-based SDK with separate STT, TTS, and LLM components. Its agent API lists at $0.04 per minute when paired with Mistral Small 4 or $0.08 per minute with other models. Realtime STT starts at $0.0035 per minute in batch and $0.008 per minute for live use, while TTS costs $13 per million characters. Self-reported latency shows full agent turns under one second and STT under 500 ms. Voxtral supports 50-plus languages for transcription and 35-plus for TTS, with EU, US, and self-host options.

Eleven Realtime integrates voice quality and low latency in one stack, while Voxtral 3 provides lower per-minute pricing and the ability to swap LLMs or run components locally. Direct head-to-head production data remains limited at launch, so teams typically run parallel pilots on both platforms to measure actual latency under their traffic patterns and token budgets.

Developers should track updates to token pricing and new transport options expected in the coming months. Testing both services on representative call volumes gives the clearest view of total monthly cost and response consistency before scaling.