What happened
Fish Audio’s Live 3 page introduces the model as part of its Voice Agents product, a platform where teams configure an agent with a prompt, voice, knowledge base and tools, then test and deploy it. The page says agents can handle inbound and outbound calls, website and app conversations, and actions against customer systems such as order lookup, refunds, calendars and CRM updates. It describes handoff to a human with call context. The company positions the product for support, bookings, orders, research and reminders.
The page says the platform supports both full-duplex and cascaded voice experiences. It does not describe the model architecture or give a measured latency figure. Fish Audio emphasizes interruption handling, silence re-engagement and response eagerness as configurable conversation behaviors. The agent can draw on uploaded PDFs, Markdown files and URLs, and can call APIs through tools and webhooks. Teams can import endpoints from cURL, add arguments and mock responses, according to the page.
Deployment options include a website widget, a React SDK, WebRTC connections from a customer’s own client, and phone numbers rented through Fish Audio or connected through a carrier. The page includes an API example for provisioning a phone number with a Fish Audio API key. It says agents can use Fish Audio’s library of more than 2 million voices across more than 30 languages, with expressive mode, speaking speed and model tier controls; voice cloning is also referenced. The page’s pricing panel says Live 3 is included at $0.035 per minute, reduced from $0.07 for a limited-time offer ending December 31, 2026. It lists a $5 starting credit without a card, phone calls at an additional $0.01 per minute, and dedicated numbers at $1.20 per month. Premium LLM use is billed at that model’s token rate.
Fish Audio also describes a testing and operations workflow. It says users can run single-turn, multi-turn and tool-call tests, with an LLM judging results against success examples, and retain version history. For deployed calls, the page lists recordings, transcripts, per-turn latency breakdowns, automatic summaries, structured data extraction and success criteria, with analysis delivered by webhook and API. The announcement page gives product capabilities and illustrative customer interactions, but not independently audited operating results.
What the archive adds
New Runtime’s earlier coverage of OpenAI’s GPT-Live-1 described full-duplex voice interaction becoming a developer-controlled API surface. That earlier note makes Live 3 a new product announcement in a known direction: Fish Audio is packaging real-time voice within a deployable agent platform, alongside tools, phone setup and operations features. An earlier note covered Fish Audio’s Drama 3 voice model, but it does not establish that the two products share a technical foundation.
Trend signal
This does not yet fit a pattern New Runtime tracks. Fish Audio’s page combines agent deployment and voice capabilities, but it does not provide enough evidence to establish a broader shift in how teams build, govern or measure agent systems.
Why now
Fish Audio’s page frames the product around handling customer calls at scale and presents a limited-time per-minute discount ending December 31, 2026. That makes the current announcement commercially time-bound, but the page does not explain why Live 3 is launching now or identify an external event driving the timing.
Why it matters
For a team considering a voice agent, the product surface matters beyond how natural a generated voice sounds. Live 3 is presented with ways to connect calls to APIs, ground answers in company materials, transfer conversations to people, and inspect recordings and transcripts afterward. Bundling those pieces may reduce the integration work required to trial a phone or web agent, especially for support and booking workflows where an agent must both converse and take an action.
The stated per-minute rate gives teams an initial input for a pilot budget, but not a complete cost model: phone calls, phone numbers and premium LLMs can add charges. The page’s testing and monitoring features also point to the work needed after a demo: defining success examples, checking tool calls, reviewing failures and tracking latency. These are useful capabilities to evaluate, not evidence that the platform will deliver reliable outcomes in production. Teams should test on their own calls and systems before treating automation claims as a basis for staffing decisions.
Limits and uncertainty
Fish Audio’s page does not provide a benchmark methodology for response latency, interruption quality, recognition accuracy or task completion. It does not report customer deployment results, call volumes, error rates or comparisons with other voice-agent platforms. The listed introductory rate is explicitly a limited-time offer, while the page does not establish what rate follows it. Security details are limited to an enterprise offering described as zero data retention; the page does not specify retention and processing terms for other plans. Builders should confirm current pricing, data handling and operational limits before putting customer conversations into production.
Skill to build now
Build a small evaluation set from the calls your team already handles: include common requests, ambiguous inputs, interruptions, failed tool calls and cases that should go to a human. For each, define the expected action and handoff condition, then run the same cases after every prompt or integration change. Fish Audio explicitly describes tests for single turns, multi-turn conversations and tool calls, making this a concrete way to assess the platform rather than relying on a polished demo.
Reader takeaway
Live 3 is presented not just as a voice model, but as a platform for configuring, connecting, testing and monitoring voice agents. Its published rate and integration options make a pilot easier to scope; they do not establish production quality or total cost. Evaluate it against real calls, tool outcomes, handoffs and the full set of charges before committing.