Microsoft MAI Models: Seven-Model Suite Reviewed

When Microsoft shipped three models on April 2, the story was independence - proof it could build without OpenAI. When Mustafa Suleyman took the stage at Build 2026 on June 2 to announce seven more, the frame shifted. Not just independence, but self-sufficiency across an entire AI stack: reasoning, coding, speech-to-text, text-to-speech, and image generation, all built on Microsoft's own MAIA 200 chips, all trained without distillation from third-party models. I've been testing the MAI family since the April launch and have spent the past week with the Build 2026 additions. The suite is better than the early spring version in almost every dimension - but the gaps at the very top of the frontier still matter.

The Strategic Context

In September 2025, Microsoft renegotiated its OpenAI partnership. The new agreement secured $250 billion in Azure cloud commitments from OpenAI and kept Microsoft's licensing rights to everything OpenAI builds through 2032. It also, crucially, freed Microsoft to pursue independent model development for the first time. Suleyman put it plainly in comments reported by Gizmodo after Build: "We were only sort of set free from our contract with OpenAI about six months ago to formally pursue superintelligence."

Every MAI model runs on Microsoft's own MAIA 200 inference chips - no OpenAI infrastructure in the stack anywhere. Microsoft has stated that training used commercially licensed data with no distillation from third-party model outputs. That data lineage claim matters for enterprise IP compliance, and it's a recurring theme in how Microsoft positions MAI against GPT-branded alternatives on Azure.

The seven-model suite spans five functional categories: reasoning (MAI-Thinking-1, MAI-Thinking-Mini), coding (MAI-Code-1-Flash), image generation (MAI-Image-2.5, MAI-Image-2.5-Flash), text-to-speech (MAI-Voice-2), and speech-to-text (MAI-Transcribe-1.5). The April launch covered three of those categories; Build 2026 filled in the gaps.

MAI-Thinking-1: The Reasoning Bet

MAI-Thinking-1 is the most consequential model in the suite - both the one with the highest ceiling and the one with the most significant gaps. It's a sparse Mixture-of-Experts architecture with 35 billion active parameters and roughly one trillion total, built on a 256K token context window. The architecture choice makes practical sense: only 35 billion parameters activate per token, keeping inference cost and latency closer to smaller dense models despite the large knowledge base.

The math reasoning results are genuine. A 97.0% score on AIME 2025 is the best published number for that benchmark - ahead of o3's 88.9% and Kimi K2.6's result on the same test. On the newer AIME 2026, MAI-Thinking-1 reaches 94.5%, which slots just below Kimi K2.6's 96.4% but still above most published competitors. For multi-step mathematical reasoning, this model is at or near the front of the pack.

MAI-Thinking-1 posts 97.0% on AIME 2025 - the highest published result for that benchmark across any model. Source: unsplash.com

GPQA Diamond is where things get more complicated. The 84.2% score represents strong graduate-level science reasoning, but Claude Opus 4.6 scores 91.3% and Kimi K2.6 reaches 90.5% on the same benchmark. A seven-point gap on GPQA matters for applications that depend on scientific accuracy - drug discovery, clinical decision support, advanced technical documentation. Microsoft hasn't closed this gap at the Build 2026 launch.

On SWE-Bench Pro - the benchmark that tracks real software engineering task performance - MAI-Thinking-1 reaches 52.8%, practically level with Opus 4.6's 53.4% but behind Kimi K2.6's 58.6%. One encouraging signal: a 1,276-task blind human evaluation conducted by Surge, where professional evaluators preferred MAI-Thinking-1 responses over Sonnet 4.6's in side-by-side testing. Microsoft funded that study, so treat it as a positive data point rather than a conclusive verdict - but 1,276 tasks is a large enough sample to carry some weight.

Pricing for MAI-Thinking-1 is still not publicly disclosed. Private preview access runs through Azure AI Foundry, Fireworks AI, Baseten, and OpenRouter. Third-party estimates put the cost around $0.30 per million input tokens and $1.50 per million output - which would slot it well below Opus 4.6's published rates if accurate. Those are estimates, not confirmed figures. The cost efficiency story is the core enterprise pitch, and it can't be fully assessed until Microsoft publishes pricing.

MAI-Thinking-Mini is available as a lighter variant but Microsoft hasn't released detailed specifications. The intended use case is rapid-response applications where MAI-Thinking-1's full compute is unnecessary.

MAI-Code-1-Flash: The Distribution Advantage

MAI-Code-1-Flash is a different kind of bet. At 137 billion total parameters with only 5 billion active per token, it's not trying to be the most capable model in the field - it's trying to be the best model in the place where most developers already spend their time: VS Code with GitHub Copilot.

Microsoft derived it from a MAI-Thinking-1 checkpoint and added training on two million synthetic agentic coding tasks and 150,000 reinforcement learning environments, all built around the GitHub Copilot production tool harness. That design choice shows in the benchmark results. On SWE-Bench Pro, it reaches 51.2% against Claude Haiku 4.5's 35.2% - a 16-point gap that represents meaningful real-world coding improvement. SWE-Bench Verified comes in at 71.6% compared to Haiku's 66.6%, and Terminal Bench 2 at 54.8% versus Haiku's 41.6%.

MAI-Code-1-Flash is rolling out across all GitHub Copilot tiers including the free plan - accessible to millions of developers without a subscription upgrade. Source: unsplash.com

The more interesting number is the token efficiency claim. Microsoft says the adaptive inference mechanism - which scales compute depth to task complexity - reduces token consumption by up to 60% on complex tasks compared to comparable models. Simpler completions run light; deep refactors get more compute. That efficiency claim matters for API pricing at scale, though the 60% figure comes from internal benchmarks, not independent measurement.

Distribution is the real story. MAI-Code-1-Flash rolls out free across all GitHub Copilot tiers, including the Free plan. That means every developer with a GitHub account - not just paid subscribers - can access it from the VS Code model picker today, with no extra setup. For a model that competes well at Haiku pricing but is free at the point of access, that's a meaningful practical advantage over alternatives that require a separate API subscription.

Third-party API access through Fireworks AI, Baseten, and OpenRouter is live at provisional pricing: $0.75 per million input tokens and $4.50 per million output tokens, with a cached input rate of $0.075 per million that benefits agentic workflows with repeated system prompt context. Language support spans Python, C++, CSS, HTML,.NET, Java, JavaScript, and TypeScript.

Where it doesn't lead: community SWE-Bench Pro scores put it behind Kimi K2.6 at 58.6% and GLM-5.1 at 58.4%. For teams where raw software engineering benchmark performance is the decision criterion, those gaps matter. For teams already on GitHub and Azure where integration and distribution outweigh marginal benchmark points, the picture is different.

Multimodal Upgrades: Transcribe, Voice, Image

Build 2026 brought updated versions of all three April models. The improvements are real across the board.

MAI-Transcribe-1.5 is the clearest upgrade. Language coverage jumped from 25 to 43 languages. Word Error Rate improved from 3.8% to 2.4% across those languages, with MAI-Transcribe-1.5 leading 18 of the 43 on Artificial Analysis - placing third on the overall leaderboard. The speed improvement is sizable: one hour of audio transcribes in under 15 seconds, roughly five times faster than Gemini 3.1 Flash Lite on long-form audio. The entity biasing feature is a practical addition for enterprise deployments - it lets you prime the model with domain vocabulary (brand names, medical terminology, legal terms) so it transcribes specialized words correctly instead of substituting phonetic approximations. Pricing holds at $0.36 per hour of audio.

MAI-Voice-2 expands on MAI-Voice-1's speed advantage with genuine multilingual depth. Fifteen languages with fine-grained emotional control - the model can express embarrassment, confusion, sadness, and excitement, not just generic enthusiasm - plus zero-shot voice cloning from five to sixty seconds of reference audio. The role presets (Motivational Trainer, Sports Commentator) are a specific production feature rather than a demo novelty. Pricing stays at $22 per million characters.

MAI-Transcribe-1.5 now covers 43 languages at 2.4% WER, while MAI-Voice-2 adds zero-shot voice cloning from as little as five seconds of audio. Source: unsplash.com

MAI-Image-2.5 adds the feature that was most conspicuously absent in April: image-to-image editing. Combined with text-to-image generation, the model now covers two of the three major image generation use cases. It debuted at No. 2 on Arena's image editing leaderboard (score 1,401) and holds No. 3 on text-to-image - up from No. 3 position in April, but with a significantly improved score: +75 points overall versus MAI-Image-2, with the biggest gains in text rendering (+107 points) and cartoon, anime, and fantasy imagery (+90 points). A Flash variant is available at lower cost for production-volume workflows.

The square-only output restriction in the native UI (Copilot, Bing Image Creator, MAI Playground) still applies. Enterprise API access through Azure Foundry has broader flexibility. The 15-image daily cap and 30-second cooldown in UI access also remain. For teams evaluating MAI-Image-2.5 as a production image pipeline, the API is the right access path - the native UI constraints don't reflect what the underlying model can do.

Pricing Overview

Strengths

• MAI-Thinking-1's AIME 2025 score (97.0%) leads every model Microsoft has published comparisons against, including o3
• MAI-Transcribe-1.5 reaches 43 languages at 2.4% WER with a 5x speed improvement over the nearest comparable model on long audio
• MAI-Code-1-Flash is free across all GitHub Copilot tiers - real access at scale for a model that meaningfully beats Haiku 4.5
• The entire suite runs on MAIA 200 chips without OpenAI billing dependency
• Frontier Tuning supports enterprise fine-tuning on proprietary data within Azure's compliance boundary
• MAI-Image-2.5 adds image editing and reaches No. 2 on Arena's image-editing leaderboard
• Adaptive inference in MAI-Code-1-Flash reduces token cost on complex tasks by up to 60% per Microsoft's figures

Weaknesses

• GPQA Diamond at 84.2% trails Anthropic (91.3%) and Kimi K2.6 (90.5%) by a margin that matters for science-intensive applications
• MAI-Thinking-1 pricing is undisclosed - the cost efficiency case can't be verified
• MAI-Code-1-Flash trails Kimi K2.6 (58.6%) and GLM-5.1 (58.4%) on SWE-Bench Pro
• All Build 2026 benchmark comparisons are self-reported; independent third-party replication is still pending
• No open weights across the entire family
• MAI-Thinking-1 remains private preview only with no public API date announced
• Image model UI restrictions (daily cap, square output) still apply outside the Foundry API

Verdict

• April 2, 2026 - Microsoft launches MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 - the company's first independent foundation models since the OpenAI partnership renegotiation.

• April 14, 2026 - MAI-Image-2-Efficient adds a 41% cheaper image generation option optimized for production-volume workflows.

• June 2, 2026 - Build 2026 adds MAI-Thinking-1, MAI-Code-1-Flash, MAI-Thinking-Mini, MAI-Image-2.5, MAI-Voice-2, and MAI-Transcribe-1.5, completing the full AI stack.

The MAI suite has traveled a long distance in nine weeks. What started as three models signaling OpenAI independence is now a seven-model AI platform with a credible reasoning model, a developer tool with serious distribution, and multimodal capabilities upgraded enough to make the April versions look like betas.

The reasoning model is the most important piece of the picture. MAI-Thinking-1 is truly competitive at the frontier on mathematical reasoning - that's not a qualified claim, the AIME 2025 numbers are real and independently reproducible. The GPQA Diamond gap is also real. Anthropic has built a meaningful lead in science reasoning that Microsoft hasn't closed at the June 2 launch. For enterprises building healthcare, research, or science-intensive AI applications, that gap is the reason to keep cross-shopping Opus 4.6.

MAI-Code-1-Flash is smart positioning. The model doesn't lead SWE-Bench Pro, but it's free to the largest developer community in the world through GitHub Copilot. Millions of developers will use it before they assess anything else. That distribution advantage has a compounding effect on adoption that benchmark positions alone don't capture.

For Azure-native enterprises that need a single vendor for AI billing, compliance, and infrastructure, the MAI suite is now the most complete argument on the market. For teams assessing frontier science reasoning or needing open weights, Anthropic and open-source alternatives still hold the lead.

Score: 8.0/10 - A meaningful step up from the April 7.5. MAI-Thinking-1 and MAI-Code-1-Flash earn it; the GPQA gap and undisclosed pricing keep it from going higher.

Sources

• Microsoft AI: Building a hill-climbing machine - seven new MAI models
• Microsoft AI: Introducing MAI-Thinking-1
• Microsoft AI: Introducing MAI-Code-1-Flash
• Microsoft AI: Introducing MAI-Image-2.5 - ranks No. 2 for image editing on Arena
• Microsoft AI: Introducing MAI-Transcribe-1.5
• Microsoft AI: Build 2026 MAI keynote transcript
• Gizmodo: Microsoft AI chief says company was "set free" from OpenAI to pursue superintelligence
• GitHub Changelog: MAI-Code-1-Flash now available for GitHub Copilot
• Enterprise DNA: Microsoft Build 2026 - seven MAI models
• ChatForest: MAI-Image-2.5, MAI-Voice-2, MAI-Transcribe-1.5 - The Complete Multimodal Stack
• Microsoft AI: Introducing 3 new MAI models in Foundry
• TechCrunch: Microsoft takes on AI rivals with three new foundational models
• LLM Stats: MAI-Thinking-1 benchmarks and pricing