Microsoft's MAI-Transcribe-2 transcribes 60 languages for $0.10/hr
Azure's new speech model tops FLEURS accuracy and runs up to 10x faster than GPT-Transcribe. Plus: Saudi Arabia opens its 428B humain-m3 to developers.
Copy markdown
$0.10 an hour, live in preview
MAI-Transcribe-2 is in public preview through Azure Speech, the MAI Playground, and OpenRouter at $0.10 per audio hour — a promo rate through Dec 31, roughly a third of what GPT-Transcribe and Gemini charge. Diarization, word-level timestamps, and automatic language detection are built in.
Fastest and most accurate, per the benchmarks
It posts a 5.2% average word-error rate to rank first on FLEURS across 60 languages and first on Artificial Analysis's accuracy-vs-latency Pareto frontier. Microsoft clocks it at 10x faster than OpenAI's GPT-Transcribe, 7x faster than ElevenLabs Scribe v2, and 5x faster than Gemini 3.5 Transcribe.
What you can ship this weekend
Cheap plus real-time makes always-on transcription worth leaving running: live meeting captions, a self-transcribing voice-note inbox, or a phone-agent pipeline. Keyword biasing, code-switching, and verbatim/clean style modes let you tune output per use case without a cleanup pass afterward.
Elsewhere: Saudi HUMAIN opens a 428B model to devs
HUMAIN debuted humain-m3 at LEAP Riyadh — a 428B-parameter MoE (23B active) on the MiniMax-M3 lineage, pretrained on 1T+ Arabic tokens. It's callable now via an OpenAI-compatible endpoint on HUMAIN Node and scored 89.37% across seven benchmarks, edging GPT-5.6 SOL and Opus 5; weights land next month under the MiniMax Community License.