Microsoft unveiled MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2—three foundational models built entirely in-house. With state-of-the-art benchmarks and aggressive pricing, the $3 trillion giant is making its independence from OpenAI clear.
Article text
Microsoft just made the clearest statement yet that it no longer needs OpenAI. On April 2, 2026, the company launched three foundational AI models built entirely in-house—speech-to-text, voice generation, and image creation—all available immediately through Microsoft Foundry.
The move isn't subtle. Mustafa Suleyman, CEO of Microsoft AI, told VentureBeat the company's mission is to be "completely independent" and achieve "AI self-sufficiency." The subtext: Microsoft is now a direct competitor to the company it once funded and relied upon.
The Three Models: What They Do
MAI-Transcribe-1 (Speech-to-Text)
This is the headline release. Microsoft claims it's the most accurate speech-to-text model available.
The numbers:
3.8% Word Error Rate
on FLEURS benchmark—besting OpenAI's Whisper-large-v3 on all 25 tested languages
Beats Google's Gemini 3.1 Flash on 22 of 25 languages
Runs on
half the GPUs
of competing state-of-the-art models
2.5x faster
than Microsoft's existing Azure Fast transcription
Technical details:
Transformer-based text decoder with bi-directional audio encoder
Handles MP3, WAV, and FLAC files up to 200MB
Built for real-world conditions: call centers, conference rooms, noisy environments
Pricing:
Starts at $0.36 per hour of audio
MAI-Voice-1 (Text-to-Speech)
A voice generation model that Microsoft is positioning against ElevenLabs and the growing voice AI startup ecosystem.
The numbers:
Generates
60 seconds of audio in one second
(60x real-time)
Supports custom voice cloning from just a few seconds of sample audio
Preserves speaker identity across generated speech
Pricing:
$22 per 1 million characters
Already powers Copilot's Audio Expressions feature, and Microsoft is now making it broadly available to developers.
MAI-Image-2 (Image Generation)
The upgraded image generation model that's been quietly running in Copilot is now available to all developers.
The numbers:
Ranked top-three on Arena.ai image generation leaderboard
2x faster generation
than its predecessor
Already in production across Bing and PowerPoint
Pricing:
$5 per 1 million input tokens, $33 per 1 million output tokens
Advertising giant WPP is already using it at scale.
Why This Is Happening Now
The backstory matters. Until October 2025, Microsoft was
contractually prohibited
from independently pursuing artificial general intelligence. The original 2019 deal with OpenAI gave Microsoft a license to OpenAI's models in exchange for building the cloud infrastructure OpenAI needed—but it came with strings attached.
When OpenAI sought to expand its compute footprint beyond Microsoft (striking deals with SoftBank and others), Microsoft renegotiated. The new terms:
Microsoft is now
free to build its own frontier models
Microsoft retains license rights to OpenAI's technology through
2032
Microsoft can pursue AGI "alone or in partnership with third parties"
Suleyman was blunt about what this means:
"Up until a few weeks ago, Microsoft was not allowed—by contract—to pursue artificial general intelligence or superintelligence independently. Since then, we've been convening the compute and the team and buying up the data that we need."
The Efficiency Story: Small Teams, Big Results
Here's what challenges the prevailing AI industry narrative: each of these models was built by
fewer than 10 engineers
.
Suleyman favors small, empowered teams over the massive research departments seen at Meta or Google. The team practices what he calls "vibe coding"—working side-by-side at round tables in a startup-like environment rather than traditional corporate desks.
The economic implications are significant:
50% less GPU cost than competitors means
lower Cost of Goods Sold
for Microsoft's own products
Smaller teams mean
lower R&D burn
compared to competitors throwing thousands of engineers at problems
Vertical integration means
Microsoft owns the entire stack
—infrastructure, models, distribution
"Humanist AI" and the Enterprise Play
Suleyman is positioning Microsoft's models as the safe choice for regulated industries. The pitch:
Clean data lineage
: Microsoft claims their training data is properly licensed, unlike some open-source models with murky copyright status
Enterprise security
: Models were "developed, tested, and rigorously evaluated for safety"
Humanist AI philosophy
: Humans remain "at the top of the food chain"
For enterprises concerned about AI compliance, this positioning matters. Microsoft is betting that Fortune 500 companies prefer Microsoft's enterprise contracts over dealing with frontier AI labs with less mature enterprise operations.
The Competitive Landscape
Microsoft is now competing on three fronts simultaneously:
Against OpenAI:
MAI-Transcribe-1 directly targets Whisper's dominance in speech-to-text
MAI-Voice-1 competes with OpenAI's voice generation
A frontier LLM to rival GPT is confirmed in development
Against Google:
Transcription benchmarks directly challenge Gemini 3.1 Flash
Image generation competes with Imagen
Both companies are building AI into their core productivity suites
Against AI startups:
MAI-Voice-1 at $22/1M characters undercuts ElevenLabs and the voice AI ecosystem
Microsoft's distribution advantage—any Foundry developer can access these models—squeezes smaller players
Against Amazon:
Aggressive pricing puts pressure on AWS AI services
Microsoft claims "best price-performance of any large cloud provider"
What This Means for Businesses
If you're building AI-powered products:
1. Evaluate Microsoft Foundry seriously
These aren't experimental models. They're production-grade, already running at scale inside Microsoft products. The quality benchmarks are real.
2. Consider the cost advantage
Half the GPU cost of competitors isn't trivial at scale. For high-volume transcription workloads, the economics could shift significantly.
3. Weigh vendor consolidation vs. specialization
Microsoft is betting you'll want one vendor for speech, voice, images, and text. Specialized players (ElevenLabs for voice, for example) may offer more features in their domain, but Microsoft offers integration convenience.
4. Watch the LLM frontier
Suleyman confirmed a Microsoft-built frontier LLM is coming. If Microsoft delivers a GPT-competitive model, the OpenAI dependency erodes further—and Microsoft's bargaining position strengthens.
The Bottom Line
Microsoft just closed its worst stock quarter since 2008. Investors are worried about AI spending and OpenAI exposure. This launch is the response: we're building our own future, controlling our own costs, and hedging against any single AI partner.