Microsoft’s MAI Models, Seven Months On: The April Trio Is Already Off the Menu


Microsoft AI infrastructure representing the MAI models launch and enterprise AI strategy

In April, Microsoft priced its first in-house transcription model at $0.36 per hour of audio. In September, its successor launched at $0.10 per hour, a figure Microsoft itself labels “a limited-time offer until the end of the year.” Between those two numbers sit five months, three generations of the same product, and the most useful lesson about Microsoft MAI models that the April launch coverage (ours included) missed.

When this post first ran on April 6, it framed MAI-Transcribe-1, MAI-Voice-1 and MAI-Image-2 as Microsoft’s hedge against OpenAI and told Azure buyers to start benchmarking. The hedge part held up. Most of the rest has changed, and a few details we published were wrong. This version corrects them and replaces the April speculation with what Microsoft actually shipped.

What we got wrong in April

Corrections first, because a news site that updates quietly is worse than one that never updates.

  • The contract date. We wrote that Microsoft renegotiated with OpenAI in September 2025. The restructured agreement was announced on October 28, 2025. That is the deal in which Microsoft says it “can now independently pursue AGI alone or in partnership with third parties,” with its IP rights extended through 2032.
  • The launch claims. We said MAI-Voice-1 generates 60 seconds of audio “in under one second” and that MAI-Image-2 debuted at number three on Arena. Microsoft’s launch post says the voice model generates “60 seconds of audio in just a single second” and that the image model debuted “as a top 3 model family” on the Arena.ai leaderboard. Both are vendor claims, and neither says what we said.
  • Three unsourced lines. We reported that MAI-Transcribe-1 ran at roughly 50% lower GPU cost than alternatives, that a 10-person team built the audio model, and that Suleyman had put frontier language models “one to two years away.” None of these appears in Microsoft’s launch material and we could not trace them to a primary source, so they are gone. So is a line about OpenAI’s revenue that carried no link.
  • “Can I use MAI models outside Azure?” We answered no. That stopped being true in June, as the next section shows.

The April figures that do check out: the 25 languages (Microsoft’s footnote defines them as the “top 25 languages by Microsoft product usage”), the claim of 2.5x the batch speed of Azure’s existing Fast transcription, and the launch prices of $0.36 per hour, $22 per million characters for voice, and $5 per million input tokens and $33 per million image output tokens for MAI-Image-2, all on the same Microsoft AI post.

Seven months, roughly a dozen models

The April trio was the opening move, not the product line. At Build on June 2, Microsoft AI announced seven models at once: MAI-Thinking-1, MAI-Code-1-Flash, MAI-Image-2.5 and a Flash variant, MAI-Transcribe-1.5, and MAI-Voice-2 with a Flash variant. The same post said the models would be “widely available for developers on OpenRouter, as well as Fireworks and Baseten.” That is the line that broke our April FAQ answer.

Since then the cadence has not slowed. Here is the lineage of the three April models, built from Microsoft’s own posts:

April model Successors What changed
MAI-Transcribe-1 (Apr 2) Transcribe-1.5 (Jun 2), Transcribe-2 (Sep 3), a streaming model (Oct 1) 25 languages to 43 to 60; $0.36/hr to a $0.10/hr promotional price through year end
MAI-Voice-1 (Apr 2) Voice-2 (Jun 2), Voice-2.1 and 2.1 Flash 23 languages; $22 per 1M characters held, Flash at $15
MAI-Image-2 (Apr 2) Image-2.5 (Jun 2), Image-2.6, Image-2.7 preview (Oct 27) 2.7 launched as editing only, “with text-to-image generation coming soon”

Two things in that table matter more than any benchmark. First, none of the April models appears on Microsoft’s current models page, which lists Transcribe-2, Voice-2, Image-2.6, Thinking-1, Code, Cyber and custom tuning. There is no deprecation notice; they simply dropped off the menu. Second, the Transcribe-2 headline price is promotional. Microsoft’s own post says it is priced at $0.10 per hour “as a limited-time offer until the end of the year,” which means a cost model built in October may be wrong in January.

A Microsoft reasoning model, which our April post treated as a distant prospect, arrived in about two months. MAI-Thinking-1 was introduced at Build and reached public preview on Foundry on August 12. Microsoft describes it as “a 35B-active, ~1T-total parameters, sparse Mixture of Experts model” with a 256k token window, and says it goes “toe-to-toe with Claude Opus 4.6 on SWE-Bench Pro” and scored 97.0% on AIME 2025. Those are Microsoft’s numbers, benchmarked against Anthropic’s previous generation, not its current one. The same post makes a claim worth noting for procurement: “We don’t distill from other labs.” Then on October 15, Microsoft AI said it had finished pre-training MAI-Base-2, “an MoE foundation model with 80 billion active and 2.8 trillion total parameters,” trained on 16,000 GB200 GPUs over 38 days. It is still in post-training, with no release date, and Microsoft says the result “will enable us to license our model weights to trusted partners through Foundry.”

The coding line is the one most enterprises already touch without knowing it. MAI-Code-1.1-Flash, released in August at “one quarter of the price of 1.0,” is described by Microsoft as “now in production in GitHub Copilot.” An October 7 update added a 3-bit build that can be downloaded and run locally on machines with more than 120GB of RAM. If your developers use Copilot, MAI is probably already in your stack; our Copilot Fusion billing checklist covers how to see which model answered which step.

The contract moved again, in the other direction

In April we wrote that Microsoft had both its own models and continued access to OpenAI’s, “a position no other company in the AI industry holds.” Three weeks later, on April 27, the two companies amended the deal again. Microsoft’s license to OpenAI IP through 2032 “will now be non-exclusive.” OpenAI “can now serve all its products to customers across any cloud provider.” Microsoft “will no longer pay a revenue share to OpenAI,” while OpenAI’s payments to Microsoft continue through 2030 “at the same percentage but subject to a total cap.” Mary Jo Foley at Directions on Microsoft read the “independent of OpenAI’s technology progress” wording as the end of the AGI clause. We covered the OpenAI side of that in what the multicloud deal means for enterprise buyers.

So the hedge now runs both ways. OpenAI no longer needs Azure to reach customers, and Microsoft no longer needs OpenAI to fill every model slot in Foundry or Copilot. MAI is less an insurance policy than a second supplier that Microsoft controls end to end, the same logic Microsoft applied to GitHub Copilot with Project Polaris at Build.

The research exit that says what MAI is optimizing for

On November 6, John Langford, who spent more than 14 years at Microsoft Research, published “The End of AI Frontiers” on his blog. AI Frontiers, the roughly 50-person group behind Autogen, was folded into Microsoft AI in September; the lab’s own page carries an “Update, Sept. 14, 2026” announcing the move. Langford’s account of what followed is blunt: “The core philosophy there is scaling, and aggressively not research,” and “I am just not comfortable being at a company where the research I do has no path to real use.” His conclusion, verbatim: “So AI Frontiers is over.” Windows Latest reported that Microsoft has not responded publicly, and notes that Langford’s claim his team’s architecture could halve training and inference cost is his own projection, not a reviewed result.

For a buyer, the useful part is not the internal politics. It is the confirmation, from someone who was inside, that MAI is run as a product shop: small teams, fast versions, measured on leaderboards and price. Microsoft says as much in its Build post (“We work in small teams with falsifiable goals over short time periods”). That is exactly what produced a new transcription model every three months. It also means the models will keep turning over faster than most enterprise change processes are built to absorb.

What to do with MAI in a Microsoft shop now

Our April advice was “start benchmarking.” That was too thin. A benchmark from April would have been obsolete twice by now. In telecom operations I have seen the version of this problem that hurts: a vendor model gets approved after a quarter of evaluation, goes into the contact center pipeline, and by the time the change ticket closes the vendor has shipped two newer versions and is quietly steering support toward them. The evaluation was sound. The approval process assumed a stable product, and the product wasn’t.

Practically, for MAI that means:

  • Approve a model family with a re-test trigger, not a version. Write the approval so that a new MAI version reruns your own eval set automatically. If you do not have one yet, our guide to building evals that do not depend on leaderboards is the place to start, because every MAI launch post leads with Arena or Artificial Analysis rankings.
  • Pin versions in Foundry and watch the models page. Microsoft removed the April models from its lineup page without a retirement notice. Treat the page as a change feed.
  • Budget at list, not promo. Model Transcribe-2 at a price you can defend after December 31, not at $0.10, until Microsoft publishes the 2027 rate.
  • Decide which route you are buying. The same MAI model reached through Foundry, OpenRouter or a local Code 1.1 Flash download comes with different data handling, logging and contract terms. Pick one deliberately, as you would for any routed model.
  • Use MAI as leverage, with numbers. Microsoft now competes with its own marketplace partners on speech and image work. A per-hour or per-character comparison against your current vendor is still the strongest thing to bring to an Azure renewal. The April post was right about that part.

The April headline called this a breakaway. Seven months later it looks more like a release train, and the buyers who do well with it will be the ones who planned for the next version before approving the current one.

Ty Sutherland

Ty Sutherland is the Chief Editor of AI Rising Trends. Living in what he believes to be the most transformative era in history, Ty is deeply captivated by the boundless potential of emerging technologies like the metaverse and artificial intelligence. He envisions a future where these innovations seamlessly enhance every facet of human existence. With a fervent desire to champion the adoption of AI for humanity's collective betterment, Ty emphasizes the urgency of integrating AI into our professional and personal spheres, cautioning against the risk of obsolescence for those who lag behind. "Airising Trends" stands as a testament to his mission, dedicated to spotlighting the latest in AI advancements and offering guidance on harnessing these tools to elevate one's life.

Recent Posts