Skip to main content
AI Choice Engine

Microsoft

Microsoft MAI-Thinking-1

Microsoft MAI-Thinking-1 is a Microsoft model with 256K context, $2 in / $8 out per million tokens, last verified October 25, 2026. Open weights: no. Public-board scores are linked to the publisher; vendor-only figures stay in the watch-out, not on the board.

MAI-Thinking-1 is Microsoft's closed text reasoning model, offered through Microsoft Foundry in public preview.

Recorded profile verification October 25, 2026Individual provider fields may changeAccess: preview — Public preview dated August 12, 2026. The previously asserted June 2 initial preview date is unverified in current primary bodies; neither a PDF filename nor model-version metadata proves availability.

AI model specification details
SpecificationMicrosoft MAI-Thinking-1
ProviderMicrosoft
TierFrontier
Context window256K
Max output64K
Input / 1M tokens$2
Output / 1M tokens$8
WeightsClosed
Parameters~1T total / 35B active (sparse MoE)
Reasoning levelsNot verifiedUnverified
Modalitiestext
ReleasedAugust 12, 2026

Pricing tiers: Azure Global Standard $2/$8 per MTok, cached input $0.20 in 32 public-region meters; two US Government regions list $2.50/$10, cached $0.25. The complete October 25 retail-feed capture has 428 items and no next page; these Thinking meters are effective August 1, 2026. Preview Global Standard only; PTU is unsupported. The current Foundry card's serialized model metadata gives exact 256,000 context and 64,000 output tokens, including reasoning. No selectable effort enum is documented.

Publisher evidence

Published benchmark results

Each result keeps its source date visible. Dates refer to source publication or capture unless the source explicitly states when it measured the result. A missing benchmark is not treated as a zero.

Published benchmark results for Microsoft MAI-Thinking-1
BenchmarkScoreSource dateSource
AIME 2025256K maximum output; vendor research, common cross-model run settings and experiment date unverifiedCross-model comparability not verified97%2026-10-02Microsoft research Table 11 (vendor-reported; October 2 capture, not experiment date) · View source
SWE-bench Verified256K total agentic context; vendor research, common cross-model run settings and experiment date unverifiedCross-model comparability not verified73.5%2026-10-02Microsoft research Table 11 (vendor-reported; October 2 capture, not experiment date) · View source
GPQA Diamond256K maximum output; vendor research, common cross-model run settings and experiment date unverifiedCross-model comparability not verified84.2%2026-10-02Microsoft research Table 11 (vendor-reported; October 2 capture, not experiment date) · View source
Terminal-Bench 2.0256K total agentic context; vendor research, common cross-model run settings and experiment date unverifiedCross-model comparability not verified46%2026-10-02Microsoft research Table 11 (vendor-reported; October 2 capture, not experiment date) · View source

Best for

  • Enterprise (Azure)
  • Reasoning
  • Microsoft ecosystem

Watch out

Preview has no SLA and Microsoft does not recommend it for production workloads. The chat model field is your deployment name, not a universal API ID; use max_completion_tokens, since max_tokens fails. Preserve the complete assistant message when returning an encrypted reasoning envelope. Tier quotas can block access, including the default low tier with zero RPM/TPM. Research Table 11 used 256K maximum output for STEM and 256K total context for agentic coding; those vendor results are not measurements of the deployed 64,000-token output contract.

Source receipts

Catalog figures for this model were checked against the following sources.

Common questions

Microsoft MAI-Thinking-1

Answered from the verified figures on this page rather than general guidance.

How much does Microsoft MAI-Thinking-1 cost per million tokens?
Microsoft MAI-Thinking-1 lists base rates of $2 per million input tokens and $8 per million output tokens. Estimate both input and output token volume and any additional charges for your workload. Azure Global Standard $2/$8 per MTok, cached input $0.20 in 32 public-region meters; two US Government regions list $2.50/$10, cached $0.25. The complete October 25 retail-feed capture has 428 items and no next page; these Thinking meters are effective August 1, 2026. Preview Global Standard only; PTU is unsupported. The current Foundry card's serialized model metadata gives exact 256,000 context and 64,000 output tokens, including reasoning. No selectable effort enum is documented.
What is Microsoft MAI-Thinking-1's context window?
Microsoft MAI-Thinking-1 accepts about 256K tokens of context. That only matters if you routinely send very long documents, large codebases, or multi-turn histories that approach that limit.
What is Microsoft MAI-Thinking-1 best for?
Microsoft MAI-Thinking-1 is a frontier tier from Microsoft. It suits enterprise (azure), reasoning, microsoft ecosystem. Preview has no SLA and Microsoft does not recommend it for production workloads. The chat model field is your deployment name, not a universal API ID; use max_completion_tokens, since max_tokens fails. Preserve the complete assistant message when returning an encrypted reasoning envelope. Tier quotas can block access, including the default low tier with zero RPM/TPM. Research Table 11 used 256K maximum output for STEM and 256K total context for agentic coding; those vendor results are not measurements of the deployed 64,000-token output contract.
Model comparison catalogBenchmark library