Selecting an AssemblyAI recognizer in October 2026 requires more than reading an old model comparison. Universal-3 Pro has been deprecated; Universal-3.5 Pro leads the prerecorded catalog, Universal-2 remains an important broad-language option, and Universal-3.6 Pro Realtime addresses live voice agents. The exact dated migration path matters to deployed applications.
The version chronology is operational, not cosmetic
AssemblyAI announced Universal-3.5 Pro Async on July 7, 2026, with 18 languages and native code switching. On July 10 it restricted legacy Universal-3 Pro for new or inactive accounts. Its August 1 migration notice explicitly states that a request pinned to `universal-3-pro` will return an error after deprecation and recommends `speech_models: [“universal-3-5-pro”]` instead. Some older model-documentation pages still describe Universal-3 Pro’s historical six-language capabilities; those pages are not a reliable current availability statement. Universal-2 is documented as supporting 99 languages. Preserve model identifiers in job logs so an integration change is visible rather than silently buried in an accuracy change.
Know the 18-language boundary before making a purchase
The July 7 first-party announcement lists English, Spanish, French, German, Italian, Portuguese, Arabic, Danish, Dutch, Finnish, Hebrew, Hindi, Japanese, Mandarin, Norwegian, Swedish, Turkish and Vietnamese for Universal-3.5 Pro Async. The August migration notice says unsupported languages may fall back to Universal-2. That fallback is useful for coverage but it changes the recognizer, so quality claims, prompts, capitalization and diarization behavior should never be applied uniformly across all 99 supported languages. Vendor benchmark results are self-reported until reproduced independently.
Prerecorded and realtime are distinct API choices
For prerecorded input AssemblyAI documented `speech_models: [“universal-3-5-pro”]` on `/v2/transcript`. On September 29, 2026 it introduced Universal-3.6 Pro Realtime for 32 languages, with streaming parameter `speech_model=universal-3-6-pro`. A streaming speech model should not be used as a drop-in spelling or latency benchmark for the async endpoint. If an application maintains two channels, pin each model independently and test input framing, finalization, interruptions and API error handling. Recheck endpoint-region availability for the exact account.
Prompting improves context but does not certify a transcript
A list of technical terms or a domain description can help with rare words, but it can also bias output toward an unspoken phrase. Create a gold-reference transcript from human-checked audio and compare raw and prompted versions. Test speaker attribution, accented names, overlapping talkers and numbers separately. Preserve the source waveform, model ID, prompt text, detected language and selected fallback so a reporter can correct quotations without guessing how a result was produced.
A safe migration decision
Inventory every call that still uses `universal-3-pro`, `universal`, legacy singular model arguments or unpinned defaults. Submit a canary set to Universal-3.5 Pro Async and keep Universal-2 available for the unsupported languages. On voice-agent applications, assess 3.6 Pro Realtime under actual interruption and silence patterns with word-level review, not only vendor speed statements. Escalate an unsupported locale or plan limit to manual handling instead of coercing a nearby language code. Separately verify billed recorded seconds and streaming connection seconds with the current first-party commercial schedule.
Worked production example
For a multilingual university archive, Universal-3.5 Pro now includes Turkish, Hindi and Japanese in its 18-language async set, while Universal-2 retains its 99-language breadth for other collections. Build a routing table that records the selected model rather than silently labeling every file as Universal-3.5 Pro. Where the provider supports fallback, confirm which model actually processed each job, and store the locale, endpoint and model result with its transcript. A terminology list for a technical lecture must not become a substitute for a corrected reference transcript.
Practical acceptance gate
Review at least fifty randomly sampled sentences in every represented language, plus intentionally difficult lists of proper nouns and numerical values. Score both the raw transcript and the human correction effort. Note whether timestamps and diarization survive a model fallback with the same interface structure. If a new language is added to a collection, test it before bulk importing the entire archive. Update comparative editorial content only when the exact product variants and source documentation are rechecked.
Region, plan and model-version notes
Documentation checked on 2026-10-25. Exact applicability: availability varies by account/endpoint and jurisdiction. Version: universal-3-5-pro async / universal-2 / universal-3-6-pro realtime; universal-3-pro deprecated. Plan: not independently verified for a specific plan. Monetary prices and quotas are not asserted unless separately verified; check the canonical AI 10 pricing ledger before purchase or publication of a tariff.
Sources & Further Reading
- Assembly Changelog — official documentation (checked 2026-10-25)
- Assembly — official documentation (checked 2026-10-25)
- Assembly Pricing — official documentation (checked 2026-10-25)
- Assembly Billing — official documentation (checked 2026-10-25)