MAI-Transcribe-2 Reality Check: Benchmarks, $0.10 Pricing and Preview Limits

Microsoft AI released MAI-Transcribe-2 on September 3, 2026 as a new batch-oriented speech-to-text model for multilingual transcription. The release matters because Microsoft is combining aggressive introductory pricing with strong independent benchmark results, but the headline claim that it is the ‘fastest, most accurate and cheapest’ speech-recognition model needs qualification.

What Microsoft actually released

MAI-Transcribe-2 is available through Microsoft Foundry and MAI Playground, and Microsoft documents it in Azure Speech as a public-preview model. The Azure integration uses the model identifier MAI-Transcribe-2 through Speech enhanced mode. The current documentation says it supports 60 languages and adds speaker diarization, word-level timestamps, keyword biasing, automatic language identification, code switching, noisy-audio robustness, and two transcription styles: verbatim, which preserves fillers and false starts, and clean, which removes fillers for a more readable transcript.

For the Azure Speech preview, Microsoft documents WAV, MP3 and FLAC inputs under 300 MB. The preview is explicitly provided without a service-level agreement and is not recommended for production workloads. That is an important practical limitation for teams considering a direct migration of mission-critical transcription systems.

Price: unusually low, but temporary

Microsoft’s launch price is $0.10 per hour of audio, offered only through the end of 2026. The company has not published the post-promotion rate in the sources checked for this analysis.

At that promotional rate, 1,000 minutes of audio costs about $1.67. Artificial Analysis independently normalizes the same model at $1.67 per 1,000 minutes on its non-streaming speech-to-text leaderboard. That is inexpensive, but it is not literally the cheapest option on the entire leaderboard: Artificial Analysis currently lists models such as StepAudio 2.5 ASR and other services at lower normalized prices. The most useful reading of Microsoft’s pricing claim is therefore that MAI-Transcribe-2 is highly price-competitive among strong managed transcription systems, not that no cheaper speech model exists.

Independent benchmark: 2.0% AA-WER and 410.7x real-time

Artificial Analysis currently measures MAI-Transcribe-2 at 2.0% AA-WER v2, a 410.7x median speed factor, and $1.67 per 1,000 minutes.

Those figures are meaningful because Artificial Analysis is not Microsoft. Its AA-WER v2 index combines about eight hours of audio from three datasets: AA-AgentTalk at 50% weight, VoxPopuli-Cleaned-AA at 25%, and Earnings22-Cleaned-AA at 25%. The speed factor is the number of seconds of input audio transcribed per second of processing time; Artificial Analysis says the displayed speed is a rolling median from recent benchmark trials using 10-minute audio.

On the current leaderboard, MAI-Transcribe-2 is second by AA-WER, behind Fun-Realtime-ASR-preview at 1.7%. It is also not the raw speed leader: Nova-3 is currently faster at 587.9x real-time. This is why a Pareto-frontier claim is more defensible than a universal ‘fastest and most accurate’ crown. MAI-Transcribe-2 combines unusually low error with unusually high throughput, even though other systems lead individual columns.

For comparison on the same Artificial Analysis table, Scribe v2 is listed at 2.2% AA-WER and 53.6x speed, Gemini 3.5 Transcribe at 2.6% and 89.9x, and OpenAI GPT Transcribe at 3.3% and 37.8x. These values can move because the speed figures are rolling measurements, so they should be treated as a dated snapshot rather than immutable specifications.

Microsoft’s FLEURS result is a different evaluation

Microsoft separately reports 5.2% average word error rate across 60 languages on FLEURS, where it says MAI-Transcribe-2 ranks first among the systems it evaluated.

This result should not be merged numerically with Artificial Analysis’ 2.0% AA-WER. They are different benchmarks, with different data and evaluation setups.

FLEURS was introduced by Google researchers as a 102-language parallel speech benchmark built from the FLoRes machine-translation benchmark, with roughly 12 hours of speech supervision per language. Microsoft is reporting a 60-language subset corresponding to the languages supported by MAI-Transcribe-2. FLEURS is useful for broad multilingual comparison, but it is not a substitute for evaluating a deployment’s own meetings, call-center audio, overlapping speech, specialist vocabulary, microphones, noise, accents, and code-switching patterns.

Microsoft’s FLEURS score is also a vendor-run result in the release material reviewed here. Artificial Analysis provides independent evidence for non-streaming accuracy and speed tradeoffs, but it does not independently reproduce Microsoft’s full 60-language FLEURS table on the same page.

Features that matter beyond WER

A low word-error rate does not automatically mean a transcript product is ready for every workflow. MAI-Transcribe-2 includes features that can materially affect usability:

• Speaker diarization attributes segments to different speakers.
• Word timestamps return offsets and durations for individual words.
• Keyword biasing lets developers provide difficult names, terminology or abbreviations as hints; Microsoft notes that hints do not force the model to output a word.
• Automatic language identification allows multilingual operation without always specifying a language in advance.
• Code switching is designed for mixed-language speech such as Hinglish and Spanglish.
• Verbatim versus clean output lets compliance-oriented workflows preserve disfluencies while note-taking products can remove fillers.

These capabilities need their own testing. A model can score well on WER while still making diarization errors, mishandling overlapping speech, assigning timestamps poorly, or degrading on the vocabulary that matters to a particular organization.

Preview status is the biggest deployment caveat

The Azure Speech documentation labels MAI-Transcribe as public preview, with no SLA, and explicitly says preview features are not recommended for production workloads. That matters more to many enterprise buyers than a small leaderboard difference.

A realistic evaluation should therefore measure at least: WER on representative audio, diarization error, timestamp quality, latency by recording length, failure rates, supported-region availability, keyword-biasing behavior, multilingual/code-switch performance, and end-to-end cost after retries and downstream cleanup.

The promotional price also ends after December 31, 2026. Until Microsoft publishes the long-term rate, teams should avoid building a multi-year cost model around $0.10 per hour.

SWE-bench and coding-agent scores are not applicable

MAI-Transcribe-2 is a specialized automatic-speech-recognition model, not a coding or general reasoning agent. SWE-bench Verified, SWE-bench Pro, Terminal-Bench, coding-agent pass rates, token context windows, and reasoning-token metrics are not comparable to ASR WER or transcription speed. Assigning those benchmarks to this model would create a false comparison.

For the same reason, ‘context window’ in the LLM sense is not the right specification to use here. The practical input constraints are the documented audio/API limits and deployment behavior.

Public feedback: evidence is still thin

Microsoft AI posted the launch announcement on X on September 3, 2026 from @MicrosoftAI. That post is vendor promotion, not independent user feedback. During this verification pass, direct X search/fetch access did not provide reliable first-hand developer review threads, and searches for attributable Reddit or forum testing did not surface enough evidence to support a sentiment claim.

Accordingly, there is no defensible public-consensus conclusion yet about MAI-Transcribe-2’s real-world quality. Early articles largely repeat the launch metrics or the Artificial Analysis leaderboard. The next useful evidence would be reproducible tests on multilingual spontaneous speech, overlapping speakers, difficult accents, domain terminology, long recordings, and diarization—not more retellings of the launch headline.

Bottom line

MAI-Transcribe-2 has a strong initial evidence profile: Microsoft reports a 5.2% average WER across 60 FLEURS languages, while Artificial Analysis independently measures 2.0% AA-WER with a 410.7x median speed factor and a normalized promotional cost of $1.67 per 1,000 minutes.

But the precise conclusion is narrower than Microsoft’s headline. It is currently one of the strongest accuracy-speed-cost combinations on the non-streaming leaderboard, not the absolute leader in every individual metric. It is also still a public-preview service without an SLA, and its $0.10/hour price is temporary. For production buyers, representative in-domain testing and the final 2027 price matter as much as the launch benchmark.