ElevenLabs on Tuesday unveiled Eleven Realtime, releasing two managed voice-agent models — Mini and Max — that bundle speech recognition, reasoning, turn-taking, and voice generation into a single package. Independent testing by Artificial Analysis found Max scored 98.8% on a benchmark of 1,000 spoken reasoning questions, compared with 75.9% for Mini. The launch marks a shift in how voice agents are built, with ElevenLabs now controlling the entire orchestration layer, including turn detection, interruption handling, and coordination between components.
The performance gap between the two models widened further in tool use, where Max completed 87.1% of tasks while Mini managed only 60.3%. ElevenLabs suggests Max for agents that execute frequent or complicated tool calls, particularly those handling bookings, customer accounts, or other multi-step processes. On Full-Duplex-Bench, which evaluates pauses, interruptions, and turn-taking, Mini scored slightly higher at 95.8% versus Max's 94.1%. Latency figures showed the models running close together, with Max averaging 570 milliseconds to first audio and Mini clocking 620 milliseconds.
According to ElevenLabs, Mini serves as the default option and can handle most use cases despite its lower benchmark scores. The company notes that teams using the Realtime API operate separately from the ElevenAgents platform, meaning features like phone and WhatsApp channels, knowledge bases, and monitoring aren't available. Migration from OpenAI's Realtime API is technically possible since Eleven Realtime uses the same protocol over WebSocket, but the company warns that WebRTC applications must change transports first and developers will need to capture and play 24kHz PCM16 audio themselves rather than relying on the SDK.
The managed approach reduces the work required to build and maintain a voice-agent stack, but it restricts developers' ability to independently troubleshoot, optimize, or swap out underlying components. Developers can still configure prompts, voices, tools, knowledge bases, and turn-taking settings through ElevenAgents, but they can't select or independently configure the models powering those features. Max costs more per token than Mini, yet its stronger performance on tool use could make it more economical for workloads where Mini needs multiple attempts to complete a task. The report notes that Mini and Max agents can transfer calls between each other, but neither can hand off to agents running in Custom mode, which lets developers select their own language and speech models. Mini also lacks Zero Retention Mode and HIPAA configurations, while Max supports both in US and EU data centers but isn't available in India or Singapore environments. The benchmarks look promising, but they don't reveal how Realtime performs under production load, and it remains unclear how much diagnostic information the platform surfaces when calls fail. The company recommends that teams needing platform features like phone integration or knowledge bases should build on ElevenAgents from the start, since starting with the bare API to move quickly may force them to rebuild those capabilities later. Voice agents built on managed infrastructure offer speed to market but require trust in the provider's ability to diagnose failures and maintain performance at scale. Organizations evaluating this architecture will need to weigh the convenience of turnkey orchestration against the loss of granular control when troubleshooting becomes necessary.