What is MAI-Thinking-1?
MAI-Thinking-1 is the first reasoning model from Microsoft AI, the in-house lab led by Mustafa Suleyman. It was announced at Build on 2 June 2026 in private preview and reached public preview on Microsoft Foundry on 12 August 2026. Microsoft says it was trained from scratch on its own infrastructure, without distillation from third-party models.
It is a sparse mixture-of-experts model with about 1 trillion total parameters, of which 35 billion are active for each token, and a 256K-token context window, which Microsoft says fits a 600-page document. It supports function calling and developer instructions through the Chat Completions API.
Microsoft's Foundry announcement lists pricing from $2 per million input tokens and $8 per million output tokens. The weights are not released.
Is MAI-Thinking-1 worth it?
MAI-Thinking-1 matters more as a statement than as a model. It is the first reasoning model Microsoft built without OpenAI, and the benchmarks Microsoft chose place it level with Anthropic's previous generation, not the current one.
Microsoft's technical report gives 52.8% on SWE-Bench Pro, against 53.4% for Claude Opus 4.6 in the same table, and 97.0% on AIME 2025 and 94.5% on AIME 2026. Coverage of the report adds 73.5% on SWE-bench Verified, 46.0% on Terminal-Bench 2.0 and 87.7% on LiveCodeBench v6. In a blind test of 1,276 tasks run by the rating firm Surge, human raters preferred it to Claude Sonnet 4.6. All of these are self-reported, and independent replication is thin.
The price is the better argument. At $2 per million input tokens and $8 per million output, it costs a fifth of GPT-6 Astra or Claude Fable 5.1 on input and less than a sixth on output, both of which sit at $10 and $50 on their makers' APIs. With 35 billion active parameters, it is also cheaper for Microsoft to serve than its 1-trillion size suggests.
The limits are the usual ones for a preview. Microsoft's materials describe text in and text out only, with no mention of image input. Microsoft publishes no maximum output length, region list or rate limits outside the Foundry model card, and preview models carry no service-level agreement. For now it is reachable only through Foundry and the MAI Playground.
Data handling follows Azure's terms for models sold by Azure: prompts are not used to train foundation models, though Microsoft notes that preview models may follow different privacy practices.
Where it earns its place
A platform team at a European insurer standardised on Azure
Run a cheaper reasoning model for internal claims triage agents
Deploy MAI-Thinking-1 in Foundry alongside a frontier model and route routine cases to it; at $2/$8 per million tokens, 100 million input and 20 million output tokens a month cost about $360.
An engineering lead evaluating coding agents
Test whether an in-house Microsoft model can replace a pricier model on routine fixes
Run the team's own pull-request tasks through it in Foundry and compare pass rates, since its 52.8% on SWE-Bench Pro is Microsoft's own figure and Terminal-Bench 2.0 results at 46.0% sit near the bottom of the frontier pack.
A maths education start-up
Generate and check worked solutions to competition problems
Use its reported 97.0% on AIME 2025 for step-by-step solutions, and spot-check outputs, since the score has not been independently reproduced.
Before you commit
- Setup
- Deploy from the Foundry model catalog into a Foundry resource in a supported region, then call it through the Chat Completions API with an API key or Entra ID.
- Integrations
- Chat Completions API with function calling and developer instructions, Foundry Agent Service, Foundry evaluation and observability tools, and the MAI Playground for trials.
- Data and privacy
- Under Azure's terms for models sold by Azure, prompts and outputs are not shared with model providers or used to train foundation models; Microsoft notes preview models may follow different privacy practices.
- Cost at scale
- From $2 per million input and $8 per million output tokens; a workload of 1 billion input and 200 million output tokens a month comes to about $3,600.
- Limits
- 256K-token context; public preview without an SLA; regions and quotas are listed only in the Foundry model card; no open weights or fine-tuning.
What’s good
- $2/$8 per million tokens, well below frontier reasoning models
- 52.8% on SWE-Bench Pro and 97.0% on AIME 2025, by Microsoft's count
- 256K-token context with function calling
- Azure data terms, identity and networking through Foundry
What’s not
- Self-reported benchmarks sit at the level of Claude Opus 4.6, a generation behind
- Text in and out only, by Microsoft's description
- Public preview with no SLA; Foundry is the only route
- No published output limit or rate limits outside the model card
Best for
Azure teams that want a cheaper reasoning model for maths, coding and agent work under Microsoft's own data terms, and can tolerate preview status.
MAI-Thinking-1 alternatives
MAI-Thinking-1: common questions
- Is MAI-Thinking-1 worth it?
- MAI-Thinking-1 is a credible first model from Microsoft's own lab, priced well for its class, but it trails the models Foundry already sells beside it. We rate it 3.7 out of 5.
- How much does MAI-Thinking-1 cost?
- Microsoft Foundry: from $2 per 1M input tokens and $8 per 1M output tokens, billed through an Azure subscription; Microsoft AI's own model page publishes no prices.
- What are the downsides of MAI-Thinking-1?
- Self-reported benchmarks sit at the level of Claude Opus 4.6, a generation behind; text in and out only, by Microsoft's description; public preview with no SLA; Foundry is the only route; no published output limit or rate limits outside the model card.
- Who is MAI-Thinking-1 best for?
- Azure teams that want a cheaper reasoning model for maths, coding and agent work under Microsoft's own data terms, and can tolerate preview status.
- How much does MAI-Thinking-1 cost for a team or in production?
- From $2 per million input and $8 per million output tokens; a workload of 1 billion input and 200 million output tokens a month comes to about $3,600.
- Is MAI-Thinking-1 safe for company data?
- Under Azure's terms for models sold by Azure, prompts and outputs are not shared with model providers or used to train foundation models; Microsoft notes preview models may follow different privacy practices.
- What are the best MAI-Thinking-1 alternatives?
- The closest alternatives we have reviewed are GPT-6 Astra, Claude Fable 5.1, DeepSeek-V4-Pro and Gemini 4 Argon.
Answers checked Nov 6, 2026.