Fish Audio introduced Live 3 on November 10, 2026: a voice model that listens while speaking and distinguishes acknowledgments from interruptions. It starts in English, through platform, API and enterprise on-premises deployment. This is not an announcement of freely downloadable weights.
Conversation and business actions require separate checks
Our editorial assessment is that natural timing solves only part of the problem. A caller may change a time or quantity midway through a sentence. Waiting correctly does not establish that the system wrote the final value to the correct order. The resulting action needs separate evidence.
Our illustrative test would compare a conversation with the resulting reservation: which date was confirmed, whether the earlier selection was cancelled and whether only one request exists. This is a proposed method, not a result tested by SIExpo. The useful metric is a correctly completed operation, rather than a pleasant voice or a quick response opening.
What is actually published
Official documentation describes Agents as public beta. It distinguishes an automatically saved draft from an immutable numbered published version; live conversations use the published configuration. Preview conversations, tests and records of tool calls are described. That provides a way to identify which configuration participated in a particular interaction.
Our conclusion for teams is to preserve the connection between a test and a version. If rules are edited after a successful test, the previous success does not establish the behavior of the new draft. The same reasoning applies when a knowledge source or connected business system changes. Every check needs a defined subject, rather than a general label saying the agent has been tested.
Language and pricing need precise reading
The platform page lists a temporary $0.035 per-minute rate through December 31, 2026. Premium language-model reasoning and telephone connectivity are billed separately. Broad marketing claims about a multilingual voice library do not automatically extend the English-language scope of Live 3 itself.
For a Bulgarian service, we would test real names, addresses, numbers and corrections in a separate language pilot before promising support. No such test was performed here. A cost assessment should include the complete conversation and associated services, rather than multiplying only one prominent rate.
Useful acceptance records would connect what the caller finally requested with the action taken and any subsequent correction. They should also distinguish a conversation that reached a human from one completed by the system. Otherwise, a smooth handover can be misreported as autonomous success, while the actual work remains elsewhere. This is our suggested accounting method, not a measured property of Live 3.
The announcement changes how the model manages interaction. Its practical value will depend on understanding, execution and the ability for a person to take over. SIExpo checked the descriptions and documentation without making paid calls or experimentally validating the company's demonstrations.
Sources: Fish Audio — Introducing Live 3, Fish Audio — voice agent platform and current pricing, Fish Audio — Agents overview and version lifecycle. The cover is a conceptual AI illustration, not a documentary photograph.
Editorial verification: 2026-11-11T17:08:13.509948+00:00.
