A caller says, “Don't change the address yet—check whether I can change it.” A voice support system must listen, wait, look up the order and avoid taking premature action. Eleven Realtime, launched by ElevenLabs on October 20, aims to take over more of that configuration work: developers choose Mini or Max while the service coordinates recognition, the language model, turn-taking and speech.
That makes the release relevant to support teams, tutors and interactive-character developers. It does not mean choosing a voice is enough for a safe launch. The useful question is what the managed setup removes—and which separately adjustable controls an existing application would give up.
You are selecting a managed stack, not just a narration model
Eleven Realtime v1 Mini targets natural conversation, tutoring and character interactions. v1 Max targets more complex instructions and multi-step tool tasks. This differs from TTS that reads supplied text and from a service that only transcribes recordings. The official compatibility guide says both versions still use a cascaded architecture: they are not fused native-audio models. The individual underlying models are not disclosed.
ElevenAgents Custom mode lets developers choose components. Switching to Realtime makes ElevenLabs responsible for certain settings, so stored LLM and TTS model choices no longer take effect. System prompts, voices, tools, knowledge bases and conversation-flow settings remain configurable. This separates control of agent behavior from selection of each underlying model; it does not remove every control.
Our assessment is that a team without much voice-engineering experience may find it easier to begin evaluating the system. Teams with custom models, a required TTS model or intricate workflows should audit those dependencies first. Fewer configuration fields do not automatically mean an easier migration. This is analysis of the product structure, not a production performance test.
Check existing handoffs before changing a support agent
Existing agents do not automatically switch because of the announcement. In an authorized ElevenAgents workspace with access, users can change Voice mode to Realtime on the Agent tab, select a model and publish. If the setting is absent, official documentation says the workspace may not yet have received the gradual rollout. A product announcement is not a promise of immediate access for every account.
Handoffs are particularly important: Realtime agents cannot transfer to or from Custom-mode agents, although Mini and Max agents can transfer between one another. A support flow that greets callers in one agent and hands after-sales questions to another could break if only one stage moves. Test the complete transfer path, not merely the opening greeting.
Agents depending on a custom LLM, Amazon Bedrock or workflow overrides of managed settings should also check the documented constraints and retain Custom where necessary. Max requires a token limit of at least 1,024, or no limit. This is a configuration requirement, not a recommendation that every response must generate that many tokens.
Protocol compatibility does not bring the whole support platform with it
ElevenLabs also offers a separate Realtime API and states that it follows OpenAI Realtime's WebSocket protocol. Teams with that kind of client can evaluate protocol reuse, but “compatible” does not mean every deployment can migrate unchanged. At launch, the endpoint supports WebSocket, not WebRTC. Browser audio capture, playback and connection handling need review for the actual transport.
There is a product boundary too. The standalone Realtime API runs separately from ElevenAgents. An API session does not automatically inherit the platform's phone or WhatsApp channels, knowledge base or monitoring. Teams needing those capabilities should evaluate ElevenAgents first instead of starting with the thinnest interface and later discovering that an entire surrounding system is missing.
At the official product entry, decide whether you need the platform or a direct API integration before comparing features. For browser connections, keep the long-lived API key on the backend and issue only a short-lived client credential using the documented flow.
A price per minute is not a price per resolved request
Checked on October 21, the official Realtime API table lists additional connected audio at $0.080 per minute for both models, plus tokens. Per million input/output tokens, Mini is $0.20/$6.00 and Max is $1.80/$9.00. A TTS price per thousand characters is not an interchangeable budget for a live conversation.
The ElevenAgents route uses the platform's per-minute rate plus model token charges. Included-minute figures are estimates affected by conversation and tool usage; exceeding concurrency limits may also incur burst pricing. Check the current pricing page and the relevant workspace bill rather than treating displayed minutes as unconditional entitlements.
For each evaluation, record connected time, tokens, tool errors and whether the request was resolved. Mini's lower token prices do not prove a lower total cost for difficult business tasks. Max's positioning does not prove greater reliability in your workflow either. Both conclusions need evidence from your own tasks.
Validate one bounded use case before promising autonomous support
Limit an initial trial to a low-risk query, such as a read-only lookup of a fictional order. Include interruptions, corrections and explicit refusals to act. Confirm that “don't change the address yet” produces no write operation. Test tool failures, long waits and human handoff too; fluent speech alone is not acceptance.
Data residency and compliance modes require separate checks. Official documentation lists Mini as unavailable with Zero Retention Mode and HIPAA configurations. Max supports those modes in US and EU environments but is unavailable in the India and Singapore data residency environments. These are service-environment and feature combinations, not a nationality-based availability rule or a guarantee that enabling a setting makes an entire business compliant.
For narration, start with the site's ElevenLabs tool overview. For subtitles from existing recordings, the local Whisper SRT tutorial addresses a different task. The meaningful comparison for Eleven Realtime is whether it reduces the work of building and maintaining live voice interactions while preserving business boundaries—not simply whether another voice sounds appealing.