AssemblyAI sets Universal-3.5 Pro as its default speech model Sept. 2, retires Universal-3 Pro; changelog names no insurance terms

Speech-to-text vendor AssemblyAI said in an August 1 changelog entry that its newest model, Universal-3.5 Pro, will become the default for customers who do not explicitly select a model, effective September 2. "Universal-3.5 Pro is now our flagship speech model, with the highest accuracy we've released to date across 18 languages," the company said. "Starting September 2, 2026, we're updating async model routing so accounts that don't pin a model get it automatically, and we're retiring the Universal-3 Pro models."

The change affects existing API integrations differently depending on how a customer's code specifies a model. AssemblyAI's changelog states that requests using the plural "speech_models" parameter set to "universal" will be "Upgraded to Universal-3.5 Pro," while requests that explicitly name the deprecated model will fail outright: "\"speech_models\": [\"universal-3-pro\"] Returns an error — Universal-3 Pro is deprecated and replaced by Universal-3.5 Pro." Real-time customers pinned to the old realtime model are handled differently: "u3-pro-rt (Universal-3 Pro Realtime) Redirected to universal-3-5-pro — no action required."

AssemblyAI's product line already includes a stated runtime-limiting control: its platform documentation, cited separately in AI Dispatch's index, describes a guardrails feature that can "redact PII and moderate content inline on audio and transcripts." The August 1 changelog entry on the model transition itself does not repeat that claim and does not address audit-trail, compliance-certification, outcome-based-payment or insurance and indemnity terms; AI Dispatch's review of the entry found no reference to any of those terms.

The deprecation gives customers roughly a month's notice: the change was announced August 1 for a September 2 cutover, according to the dated changelog entry.