MAI-Thinking-1 Review 2026: Microsoft Ditches OpenAI With Its First Reasoning Model

Microsoft has officially stopped being an OpenAI reseller dressed up as an AI company. On June 2, 2026 — day one of Microsoft Build 2026 — Mustafa Suleyman’s Microsoft AI team unveiled MAI-Thinking-1, the company’s first in-house reasoning model, alongside six sibling models including the inference-efficient coding workhorse MAI-Code-1-Flash. None of them were distilled from GPT-class teacher models. None of them were trained on OpenAI data. And the headline number — 94.5% on AIME 2026 — is a frontier result for a 35B-active-parameter model.

We have spent the past 24 hours putting MAI-Thinking-1 through its paces on the Microsoft Foundry private preview, running it head-to-head against Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5. This review covers what is genuinely new about the architecture, how the benchmarks hold up on real workloads, where MAI-Thinking-1 quietly wins, and the political subtext you should not ignore.

The Short Version

If you have ten seconds:

• First Microsoft-native frontier model. No OpenAI distillation, no third-party teacher, trained entirely on commercially licensed enterprise data inside Microsoft’s own infrastructure.
• Architecture: sparse Mixture of Experts, ~1T total parameters, 35B active per token, 256K context window, function calling and multi-step instruction-following built in.
• Benchmarks: 97.0% on AIME 2025, 94.5% on AIME 2026, matches Claude Opus 4.6 on SWE-Bench Pro, and was preferred over Claude Sonnet 4.6 in Surge’s blind side-by-side ratings.
• Sibling model MAI-Code-1-Flash hits 85.8% on Microsoft’s adversarial coding benchmark and ~51% on SWE-Bench Pro while using up to 60% fewer tokens than comparable models, and ships into every paid GitHub Copilot tier this week.
• Availability: private preview through Microsoft Foundry, compatible with the OpenAI Chat Completions API drop-in pattern. General availability expected in the second half of 2026.
• Why it matters: Microsoft now has a credible non-OpenAI option for Copilot, Azure AI Foundry, and Windows Copilot+ devices — and the leverage that comes with it.

If you have already built around OpenAI’s API on Azure, MAI-Thinking-1 is a near-zero-friction A/B test you should run this month. If you do not care about model politics, it is still the strongest reasoning-tier model you can get on Microsoft Foundry today.

What Is MAI-Thinking-1?

MAI-Thinking-1 is Microsoft’s first reasoning-class large language model — the same category as OpenAI’s o-series, DeepSeek-R1, and Anthropic’s extended-thinking Claude variants. Like those models, it spends time on an internal chain of thought before emitting a final answer, and it scales test-time compute by adjusting how long it thinks.

It sits at the top of seven new MAI models announced at Build 2026, a group that includes:

• MAI-Thinking-1 — frontier reasoning and agentic flagship.
• MAI-Code-1-Flash — small, fast coding model now powering free-tier GitHub Copilot.
• MAI-Voice-2 — next-gen text-to-speech for Copilot Voice and Windows narration.
• MAI-Image-2 — text-to-image, billed as the enterprise alternative to DALL·E 3 and Stable Diffusion.
• MAI-Embed-1 — long-context embedding model for Foundry retrieval and Microsoft 365 Work IQ.
• MAI-OCR-1 and MAI-Translate-1 — utility models for document workflows.

The strategic point is that Microsoft now controls the whole stack — pretraining data, base model, fine-tuning recipe, RLHF pipeline, inference serving, and the application surface. That is the configuration OpenAI, Anthropic, and Google have always enjoyed and the one Microsoft, despite owning Azure and Copilot, never fully had.

Architecture: A Real Mixture of Experts, Not a Marketing One

The technical claim worth focusing on is the 35B active / ~1T total parameters split. That is a real sparse Mixture of Experts in the style of DeepSeek V4 and Mistral’s frontier line — not the “wide-and-shallow” MoE that some vendors have been calling sparse to inflate parameter counts.

What that gets you in practice:

• Cheaper inference per token than a dense 100B-class model with similar quality. Microsoft has not published its serving cost, but the activation ratio (3.5%) is in the same range as DeepSeek V4 and Mistral Vibe.
• Faster first-token latency, especially on Foundry’s H200-class hardware in the East US 2 and Sweden Central regions where the private preview is concentrated.
• Better long-context behavior. The 256K context window holds up well — we ran an 80K-token codebase summarization that Opus 4.8 also handles, and MAI-Thinking-1 matched it on factual recall while being noticeably faster.

The model supports function calling, multi-layered instruction following, structured outputs (JSON Schema), and is exposed through the OpenAI-compatible Chat Completions API that Foundry has used since 2024. If your code is built around openai.ChatCompletion.create(...) against an Azure endpoint, swapping gpt-5.5 for mai-thinking-1 is a one-line change.

Benchmarks: What Holds Up

Microsoft published a long benchmark table at Build. Here is the subset that matters, with our own quick replications where we could run them.

Benchmark	MAI-Thinking-1	Claude Opus 4.8	GPT-5.5	Gemini 3.1 Pro
AIME 2025	97.0%	96.7% (USAMO 2026 proxy)	94.1%	95.3%
AIME 2026	94.5%	not reported	90.8%	91.6%
SWE-Bench Verified	86.4%	88.6%	82.1%	78.9%
SWE-Bench Pro	64.7% (matches Opus 4.6)	69.2%	58.6%	54.0%
GPQA Diamond	91.2%	90.7%	89.4%	88.0%
MMLU-Pro	88.9%	88.4%	87.1%	86.5%
Function-calling reliability (BFCL v2)	95.1%	94.7%	93.0%	91.8%

A few honest observations:

1. The math result is real. 94.5% on AIME 2026 is a frontier number, and the model genuinely shows its work — long, well-structured chains of thought that you can read and check.
2. Coding is competitive, not dominant. Opus 4.8 still wins SWE-Bench Pro by ~4.5 points, and Claude Code plus Dynamic Workflows still beats anything Microsoft has shipped at multi-day agentic refactors.
3. The “preferred over Claude Sonnet 4.6” claim from Surge is plausible, but Sonnet 4.6 is the mid-tier model. The fair comparison is against Opus 4.8, and Opus still feels more thoughtful on open-ended judgment calls (writing performance reviews, balancing PR feedback, etc.).
4. Function-calling reliability is the sleeper win. 95.1% on BFCL v2 is the highest result we have measured. If you are building agents, this matters more than another point of SWE-Bench.

What We Tested

We ran MAI-Thinking-1 on five workloads we use to evaluate every new frontier model:

1. Long-horizon math (AIME 2026 + IMO 2024 shortlist)

We pulled fifteen problems from outside Microsoft’s reported benchmark set. MAI-Thinking-1 nailed 13/15 with clean working. Opus 4.8 also got 13/15. GPT-5.5 got 11/15. This is genuinely a tie at the frontier, which is itself a result given how far Microsoft was behind 18 months ago.

2. Codebase comprehension (80K-token monorepo Q&A)

We loaded an 80K-token TypeScript codebase and asked twelve cross-cutting questions (“which modules touch the auth cookie”, “where does the rate limiter break ties between IPs”, etc.). MAI-Thinking-1: 11/12 correct, with citations. Opus 4.8: 12/12. The one miss for MAI was a subtle dependency injection question that Opus also took two thinking passes to get.

3. Refactor + tests (single-file Python → Rust port)

A ~600 line algorithmic Python module. MAI-Thinking-1 produced a working Rust translation with cargo test passing on first run after one round of compile-error feedback. Opus 4.8 in standard mode did the same. Where Opus pulled ahead was on idiomatic Rust — borrows, lifetimes, and avoiding unnecessary clone(). MAI’s output worked but read like Rust written by a senior Python engineer.

4. Agentic data pipeline (multi-tool, multi-step)

We gave both models the same agent loop with five tools: SQL, HTTP, file I/O, a calculator, and a “ask user” escape hatch. The task: build a daily revenue snapshot from three internal sources. MAI-Thinking-1 finished in 14 tool calls, Opus 4.8 in 11. MAI made one redundant SQL query but never got stuck. This is exactly where the high BFCL score shows up — no malformed function calls, no JSON parse errors.

5. Open-ended judgment (“rewrite this product brief for a non-technical exec”)

Opus 4.8 still wins this one. MAI-Thinking-1’s output was correct, complete, and a little flat. Opus had a better sense of which paragraphs to drop entirely.

MAI-Code-1-Flash: The Sleeper Hit

The bigger immediate impact may come from MAI-Thinking-1’s smaller sibling, MAI-Code-1-Flash, which Microsoft is rolling out to every GitHub Copilot tier — Free, Pro, Pro+, and Max — this week.

Headline numbers:

• 85.8% adjusted accuracy on Microsoft’s adversarial coding benchmark.
• ~51% on SWE-Bench Pro, a frontier-class result for a small model.
• Up to 60% fewer tokens on harder coding tasks compared with similar-sized models.
• Trained inside the GitHub Copilot production harness rather than benchmarked externally and then deployed, which Microsoft says is a reliability win for agentic coding loops.

The token-efficiency claim is the interesting one. In our limited testing — porting a small Vue 2 component to Vue 3 across five files — MAI-Code-1-Flash used roughly 40% fewer output tokens than Claude Haiku 4.5 for an equivalent diff. At Copilot scale that compounds into a real margin improvement for Microsoft.

If you have been ambivalent about GitHub Copilot versus Cursor or Windsurf, the upgrade to MAI-Code-1-Flash + the existing Claude Opus 4.8 option in Copilot Max is a real reason to revisit.

Pricing and Availability

Microsoft has not published list prices for MAI-Thinking-1 yet — it is private preview on Microsoft Foundry with general availability “in the second half of 2026.” That said:

• MAI-Code-1-Flash is included at no extra cost in every GitHub Copilot tier, including Free. This is a clear shot across OpenAI’s bow.
• MAI-Thinking-1 preview access is being granted through Microsoft Foundry to existing enterprise customers via the Foundry portal. Expect pricing in the $3–6 input / $15–25 output per million tokens range based on Foundry’s pattern for previous Microsoft-served frontier models.
• Azure regions at preview: East US 2, Sweden Central, Japan East. EU compliance under the AI Act high-risk obligations (which take effect August 2026) is being handled through the Sweden Central deployment.

Compatible APIs:

• Chat Completions API (OpenAI-compatible) — one-line drop-in if you already call Foundry.
• Responses API — Foundry’s newer thinking-aware interface, recommended for reasoning workloads.
• Function calling with parallel tool calls.
• Structured outputs via JSON Schema.

Pros and Cons

Pros

• Frontier math and reasoning at a 35B active-parameter cost.
• 256K context window with strong long-context recall.
• Best-in-class function-calling reliability (95.1% BFCL v2).
• OpenAI-compatible API means zero migration cost from Foundry/Azure OpenAI.
• Strategic independence — no OpenAI lock-in, useful leverage if you negotiate Azure commits.
• MAI-Code-1-Flash bonus — every Copilot user gets the new model immediately.

Cons

• Private preview only — not yet GA, no public pricing, capacity is limited.
• Coding lags Opus 4.8 by ~4.5 points on SWE-Bench Pro.
• Open-ended writing is still a notch below Opus 4.8 and Gemini 3.1 Pro.
• Tooling ecosystem is thin — no MAI-native equivalent of Claude Code’s Dynamic Workflows or Cursor’s agent harness.
• Region availability is narrow at preview, which matters for EU and APAC compliance.
• Benchmarks are Microsoft-reported for now; expect independent reproductions to land over the next few weeks.

Who Should Use MAI-Thinking-1?

• Azure-committed enterprises already on Foundry — yes, run a parallel pilot this month. The migration cost is trivial and the reasoning quality justifies it for analytical workloads.
• Teams building agentic apps that care more about function-calling reliability than raw coding score — MAI-Thinking-1 is genuinely the best we have measured here.
• Math, science, and finance workloads that previously routed to o-series models — MAI matches or beats them at, we expect, lower cost.
• GitHub Copilot users on any tier — MAI-Code-1-Flash is rolling out automatically, you do not need to opt in.

Who Should Stick With What They Have

• Heavy Claude Code users doing long-horizon agentic refactors — Dynamic Workflows on Opus 4.8 still has no Microsoft equivalent.
• Creative writing and brand voice work — Opus 4.8 and Gemini 3.1 Pro remain ahead.
• Multimodal-heavy workflows (image and video understanding) — wait for MAI-Image-2 and the multimodal MAI-Thinking variant Microsoft has hinted at for fall 2026.

The Political Subtext

This release is the loudest signal yet that Microsoft is gradually decoupling from OpenAI. The Build 2026 keynote barely mentioned GPT, the OpenAI logo was absent from the Foundry model-card screen, and Suleyman explicitly said the MAI line was “trained without distillation from third-party models.” The Surge blind-evaluation framing — “preferred over Claude Sonnet 4.6” rather than “matches GPT-5.5” — also strategically benchmarks against Anthropic rather than the partner whose models still account for the bulk of Foundry traffic.

If you are an enterprise running on Azure, this is leverage. You should be using it at your next Microsoft renewal — having a credible second source on Foundry makes your negotiating position materially stronger.

Verdict

4.5 / 5. MAI-Thinking-1 is the first time Microsoft has shipped a model where the answer to “why use this over GPT-5.5 or Opus 4.8?” is not just “because it is on Azure.” The math results are real, the function-calling reliability is best-in-class, and the architecture is honest sparse MoE rather than dense-pretending-to-be-sparse. The 4.5-point gap on SWE-Bench Pro versus Opus 4.8 keeps it out of a perfect score, and the private preview state means you cannot rely on it in production yet — but if you are on Azure, this is the most consequential model release of the quarter.

We are recommending it for any team that already lives on Foundry and would benefit from a frontier reasoning model that they can call with the same SDK they already use. We will be watching closely for the GA release in H2 2026 and for independent reproductions of the AIME 2026 number — at which point the verdict could move up.

For comparisons against the rest of the 2026 frontier, see our Gemini 3 vs GPT-5.5 vs Claude breakdown, our Claude Opus 4.8 review, and our GPT-5.5 review. If you are picking between coding tools that can host MAI-Code-1-Flash, see our best AI coding assistants guide.