Microsoft MAI-Thinking-1 Matches Opus 4.6 on SWE-Bench Pro

Microsoft MAI-Thinking-1 Matches Opus 4.6 on SWE-Bench Pro

Microsoft MAI-Thinking-1 is a 35B active parameter sparse MoE reasoning model (approximately 1T total parameters) that matches Claude Opus 4.6 on SWE-Bench Pro at 52.8% and achieves 97% on AIME 2025. It was trained from scratch on 30 trillion tokens using 8K GB200 GPUs on Azure infrastructure, with zero distillation from third-party models and fully traceable training data. (Source: Microsoft AI technical report, June 2026)

The Real Problem

The reasoning model market in 2026 is dominated by distil-and-copy training strategies. Most new models are fine-tuned from outputs of existing frontier models, which creates three problems: unlicensed training data, opaque model capabilities that depend on the teacher model, and a consolidation of architectural approaches that reduces diversity. Enterprises looking to deploy reasoning models internally cannot audit where the training data came from.

[ STAT ] MAI-Thinking-1 achieves 52.8% on SWE-Bench Pro, matching Claude Opus 4.6, and 97% on AIME 2025. Blind human raters on Surge prefer it to Sonnet 4.6 for overall quality. (Source: Microsoft AI technical paper, June 2026)

Microsoft claims MAI-Thinking-1 is built entirely on clean, enterprise-grade data. Every training source is documented and appropriately licensed. The company positions this as the key differentiator, not benchmark scores. For enterprises that cannot use models trained on uncleared data, this is the decisive factor.

What This Actually Does

MAI-Thinking-1 is a reasoning model, not a general-purpose chat model. It is optimized for math, coding, and structured reasoning tasks where multi-step deliberation matters. The model uses a 256K context window and supports function calling, complex instruction following, and human-in-the-loop calibration.

[TOOL: MAI-Thinking-1] Handles agentic coding tasks including reading code, editing files, running tests, bug fixing, observing failures, and recovering from intermediate mistakes. It adapts to human feedback and calibrates to user intent across single and multi-turn tasks.

The model architecture is a sparse mixture of experts with 78 layers, 6,656 hidden dimension, 8/512 KV heads, and 8/80 active experts. The base model was pre-trained on 30 trillion tokens, with mid-training phases totaling 3.55 trillion additional tokens.

Who This Is Built For

For enterprise AI teams deploying reasoning models in production who need clean data provenance: MAI-Thinking-1's training dataset is fully documented and appropriately licensed. No third-party distillation, no opaque data sources. This matters for compliance, audit, and legal review.

For software engineering teams who want a model that can act as an agentic coding assistant without the cost of running a 400B+ parameter model: at 35B active parameters, MAI-Thinking-1 has a smaller inference footprint than much larger models but matches Claude Opus 4.6 on SWE-Bench Pro.

For Microsoft Foundry customers who need a reasoning model with enterprise safety guardrails, copyright protection, and multi-region deployment: the model is available in private preview on Microsoft Foundry with built-in safety features.

How It Runs: Step by Step

1. API access: MAI-Thinking-1 is available through the Chat Completions API on Microsoft Foundry, which means the migration path from existing OpenAI-compatible models is straightforward. The API endpoint, request format, and response schema match the widely used spec.

2. Model selection: Choose the model name "mai-thinking-1" in your API call. The model supports system prompts, user messages, tool definitions, and streaming responses. No special configuration is needed beyond what standard Chat Completions API calls require.

3. Multi-turn reasoning: The model's strengths emerge in multi-turn tasks. Start with a task that requires reasoning and iteration: give the model a codebase reading task, a debugging problem, or a math proof. The model will produce a reasoning trace before answering.

4. Function calling: Define tools and pass them in the tools parameter. The model can call functions, receive results, and continue its reasoning loop. This is the mechanism for agentic coding workflows.

Setup and Tools

The model is accessible via Microsoft Foundry in private preview at microsoft.ai/models/mai-thinking-1. The Chat Completions API format means existing code that calls OpenAI or Anthropic APIs can be adapted with a URL and API key change. No SDK changes required.

One gotcha: MAI-Thinking-1 is a reasoning model, which means it produces longer outputs than non-reasoning models for the same task. The reasoning trace is included in the output tokens and billed accordingly. If you need a quick factual answer without reasoning overhead, a non-reasoning model like GPT-4o or Claude Sonnet 4.6 will be cheaper and faster. Use MAI-Thinking-1 when the task genuinely benefits from step-by-step reasoning.

The Numbers

[ STAT ] SWE-Bench Pro: 52.8%, matching Claude Opus 4.6. (Source: Microsoft AI technical report, June 2026)

[ STAT ] AIME 2025: 97.0%, placing it among the strongest models on advanced math reasoning. (Source: Microsoft AI technical report, June 2026)

[ STAT ] LiveCodeBench v6: 87.7%, demonstrating strong algorithmic coding performance. (Source: Microsoft AI technical report, June 2026)

[ STAT ] Human preference: Independent raters on Surge prefer MAI-Thinking-1 to Sonnet 4.6 for overall quality in blind side-by-side evaluations across single and multi-turn tasks. (Source: Microsoft AI, June 2026)

What It Cannot Do

1. MAI-Thinking-1 is a text-only reasoning model. It does not support image, audio, or video inputs. If your workflow requires multimodal understanding, you need a separate vision model in your pipeline.

2. The model is optimized for math, coding, and reasoning. Tasks that require creative writing, long-form narrative generation, or open-ended conversation are not its strengths. Use a general-purpose model for those.

3. At 35B active parameters, MAI-Thinking-1 is cost-efficient compared to 400B+ models, but it is not small enough to run on local hardware (laptop or edge device). It requires cloud inference or on-premises GPU servers.

Start in 10 Minutes

1. (5 min) Sign up for Microsoft Foundry private preview at microsoft.ai/models/mai-thinking-1. The waitlist approval may take 1-2 weeks, so start this first.

2. (3 min) Review the API documentation. The Chat Completions format requires no SDK changes if you already call OpenAI-compatible APIs. Set the model field to "mai-thinking-1".

3. (2 min) Test with a single reasoning task: send a SWE-Bench-style prompt asking the model to fix a bug in a given code snippet. Evaluate whether the reasoning trace quality meets your expectations.

4. (Ongoing) Run your existing agentic coding workload through the model. Compare output quality and cost against your current model. MAI-Thinking-1 excels at tasks where step-by-step reasoning matters.

Frequently Asked Questions

Q: How does MAI-Thinking-1 compare to Claude Opus 4.6 on SWE-Bench? A: MAI-Thinking-1 matches Claude Opus 4.6 on SWE-Bench Pro at 52.8%. This is notable because MAI-Thinking-1 has 35B active parameters while Opus 4.6 is estimated to have significantly more parameters. The model achieves this without distillation from any third-party model, including Claude. (Source: Microsoft AI technical report, June 2026)

Q: Is MAI-Thinking-1 built on distilled data from OpenAI or Google models? A: No. Microsoft explicitly states that MAI-Thinking-1 was trained from scratch with zero distillation from other labs. The training corpus was built entirely from publicly available and acquired data sources, all appropriately licensed. This is the company's primary differentiation in the reasoning model market. (Source: Microsoft AI technical paper, June 2026)

Q: What hardware was used to train MAI-Thinking-1? A: The base model was pre-trained on 8K GB200 GPUs on a Microsoft-operated cluster within Azure. The pre-training phase used 30 trillion tokens, followed by mid-training phases totaling 3.55 trillion tokens. Microsoft's in-house distributed training infrastructure coordinated the workload across the cluster.

Q: Will MAI-Thinking-1 be open source? A: Microsoft has not announced open-source plans. The model is currently available in private preview on Microsoft Foundry as an API. The model card and technical paper are publicly available at microsoft.ai/pdf/MAI-Thinking-1-Model-Card.PDF, but the weights have not been released.

Q: Can MAI-Thinking-1 be used for agentic coding workflows? A: Yes. The model was trained for multi-turn software engineering tasks: reading code, editing files, running tests, fixing bugs, and recovering from errors. It supports function calling and complex instruction following, which enables agentic loops. Microsoft trained it using 8M+ reinforcement learning environments for software engineering tasks.