Universal-3.5 Pro
AssemblyAI
Universal-3.5 Pro is AssemblyAI's flagship pre-recorded speech-to-text model, requested as speech_models: ["universal-3-5-pro"] on /v2/transcript. It transcribes 18 languages with native mid-sentence code switching, takes contextual prompting plus keyterms prompting of up to 1,000 words, and solves transcription and speaker attribution jointly rather than with a separate diarizer. Requests in other languages fall back automatically to Universal-2, which covers 99. AssemblyAI's changelog dates the pre-recorded launch to July 7, 2026 and the separate streaming sibling to June 23, 2026.
Strengths
- 4.35% average WER across AssemblyAI's four-dataset pre-recorded suite, the leading row in its own comparison table
- Native mid-sentence code switching across 18 languages with no prompt configuration needed
- Contextual prompting plus keyterms prompting of up to 1,000 words on pre-recorded requests
- Transcription and speaker attribution solved jointly instead of bolting on a separate diarizer
- Automatic fallback to Universal-2 for unsupported languages, giving 99-language coverage without extra configuration
- Optional spoken-form output (format_text: false) that leaves numbers and dates unnormalized for verbatim workflows
- Medical Mode add-on (domain: "medical-v1") for medication, procedure, condition and dosage terminology in English, Spanish, German and French
Caveats
- Only 18 languages natively; anything else silently falls back to the cheaper, less accurate Universal-2, and the multilingual table reports language-level scores for just 5 of 20 evaluated languages.
- Diarization is weak in absolute terms at 30.17% average cpWER, reaching 48.22% on NOTSOFAR dev, and it loses to Speechmatics on AMI (27.36% vs 20.82%) and to ElevenLabs Scribe V2 on CALLHOME dev/eval (17.18% vs 14.82%).
- Entity accuracy is uneven: 43.70% missed-entity rate on money (worse than OpenAI GPT-4o Transcribe's 37.30%), 16.33% on temporal expressions (against Mistral Voxtral Mini's 6.21%) and 54.37% on URLs.
- Every published benchmark is AssemblyAI's own, on its own dataset selection including a synthetic medical set, with no third-party evaluation cited.
- Closed model, API only, and not yet the async default - AssemblyAI says routing switches to it on September 2, 2026; streaming is a separate model billed at $0.45/hr on connection time including idle time.
- Vendor
- AssemblyAI
- License
- Proprietary
- Modalities
- audiotext
| Benchmark | Score | Max |
|---|---|---|
| WER, average across AssemblyAI's pre-recorded suite | 4.3 | 100.0 |
| WER, accented English (India) | 5.2 | 100.0 |
| Diarization cpWER, average over DiPCo, CALLHOME, NOTSOFAR and AMI | 30.2 | 100.0 |
| Multilingual global WER | 7.8 | 100.0 |
| Code-switching WER, average over en paired with de/es/fr/it/pt | 7.6 | 100.0 |
Capabilities
Resources
Pricing
$0.21 per hour of audio for pre-recorded transcription, with volume discounts, against $0.15/hr for Universal-2. The separate Universal-3.5 Pro Streaming model is $0.45/hr billed on WebSocket session duration - the time the connection is open, including idle time - versus $0.15/hr for the Universal-Streaming models. New accounts receive $50 in free credits.