by Microsoft

MAI-Thinking-1 review, pricing and limits

Microsoft AI's first from-scratch, first-party reasoning model, built to power agentic coding and math-heavy workflows in Foundry and Copilot without relying on OpenAI's models.

Best for
Agentic coding inside the Microsoft ecosystem
Not for
Vision or audio-driven applications
  • Preview
  • Proprietary

Should you use MAI-Thinking-1?

MAI-Thinking-1 is Microsoft AI's first in-house reasoning model, a 35B-active, roughly 1T-total MoE with a 256K context window. Microsoft Foundry now lists it at $2 input and $8 output per 1M tokens in public preview, and Microsoft reports 73.5% on SWE-bench Verified and 97.0% on AIME 2025.

Best for

  • Agentic coding inside the Microsoft ecosystem 73.5% SWE-bench Verified and 52.8% SWE-bench Pro, described as toe-to-toe with Claude Opus 4.6 on SWE-bench Pro.
  • Math and graduate-level science reasoning 97.0% AIME 2025, 94.5% AIME 2026, and 84.2% GPQA Diamond.
  • Enterprise teams standardizing on Azure/Foundry governance First-party Microsoft model avoiding dependency on OpenAI weights, with Foundry's built-in enterprise security and data residency controls.

Not for

  • Vision or audio-driven applications Text and tool-calls only; no image, audio, or video input support.
  • Production workloads that need a general-availability SLA The model is in Foundry public preview, so rates, limits and availability can change before general availability.
  • Long, messy multi-turn terminal agent automation Terminal-Bench 2.0 score of 46.0% and MultiChallenge score of 53.0% trail top frontier agentic models.

MAI-Thinking-1 pricing

Input, per 1M tokens
$2.00/1M
Output, per 1M tokens
$8.00/1M
Cached input, per 1M tokens
$0.20/1M
Pricing model
per-token
Free tier
No

What a real job costs

JobInputOutputTotal
Summarise a 20-page PDF$0.060$0.0080$0.068
Support reply$0.0040$0.0024$0.0064
One coding agent run$0.400$0.160$0.560

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

The Global deployment is listed by Azure at $2.00 per 1M input tokens, $0.20 per 1M cached input and $8.00 per 1M output, read on October 23, 2026 while the model is in public preview. Microsoft has not published a regional-deployment rate, a batch discount or a free tier for it. Third-party trackers quote different figures; only the Azure list is treated as the source here. For scale, Microsoft's smaller MAI-Code-1.1-Flash lists at $0.20 and $1.20 in Hokai's records.

What MAI-Thinking-1 can do

MAI-Thinking-1, released June 2, 2026 by Microsoft AI, is a 256K-context reasoning model built as a sparse MoE with 35B active parameters. Microsoft Foundry prices its Global deployment at $2 per 1M input and $8 per 1M output tokens in public preview, and Microsoft reports 97.0% on AIME 2025.

What goes in and what comes out

Reads

  • text
  • tool-calls

Writes

  • text
  • tool-calls

Capabilities

  • Function callingSupported
  • Structured outputSupported
  • Tool useSupported
  • Audio inputNot supported
  • Audio outputNot supported
  • Code executionNot supported
  • Video inputNot supported
  • VisionNot supported
  • Web browsingNot supported
Reasoning modes
extended-thinking
Long context recall
unverified

Key features

  • 256K Context Window

    Handles roughly 600 pages of text in a single request, well suited to large codebases and long documents.

  • 35B-Active MoE Architecture

    Sparse Mixture of Experts with 35B active out of ~1T total parameters, built on the internal MAI-Base-1 foundation model.

  • No-Distillation Training

    Trained from scratch on commercially licensed data with zero distillation from third-party models, a first for Microsoft's flagship reasoning line.

  • Joint Safety-Capability RL Loop

    Safety rewards run through the same reinforcement learning infrastructure as capability training, targeting both unsafe compliance and unnecessary refusal as defects.

  • Function Calling and Developer Instructions

    Native tool use and Chat Completions API compatibility, including a distinct developer-instruction role.

Known limitations

  • No vision, audio, or video input; text and tool-calls only.
  • Still a preview model: Microsoft can change the $2 and $8 per 1M price before general availability.

Who it is for

  • Enterprise teams standardized on Microsoft Azure and Foundry governance
  • Agentic coding teams wanting a first-party Microsoft reasoning model
  • Teams doing math-heavy or graduate-science reasoning workloads
  • Copilot-integrated developer teams avoiding OpenAI dependency

Supported languages

  • English

MAI-Thinking-1 benchmarks

Standout result

97.0%

AIME 2025 score, Microsoft's headline math reasoning claim for MAI-Thinking-1

Source: Microsoft AI, Introducing MAI-Thinking-1, June 2, 2026

Where it sits

Priced around the middle of the 107 GA models with a published price (rank 71), and mid-pack on SWE-bench Verified (rank 24 of 40). Ranked against GA models; this record is not GA.

MeasureThis modelRankMedian
Blended price (3:1)$ per 1M tokens$3.50/M#71 / 107peer median $1.75/M
SWE-bench Verified% solved73.5%#24 / 40peer median 76.3%
GPQA Diamond% correct84.2%#40 / 64peer median 87.6%

Ranks are against GA models on Hokai that publish the same figure; ties share a rank.

Benchmark scores

  • AIME 2025 vendor-reported · 02 Jun 202697%Competition-level maths problems from the 2025 exam, % solved.
  • MMLU-Pro vendor-reported · 02 Jun 202685%A harder version of the 57-subject knowledge exam, % correct.
  • GPQA Diamond vendor-reported · 02 Jun 202684.2%PhD-level science questions that are hard to search for, % correct.
  • SWE-bench Verified vendor-reported · 02 Jun 202673.5%Real GitHub issues fixed end to end, % solved.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Strengths and limits

Strengths

  • 97.0% AIME 2025 and 84.2% GPQA Diamond, strong reasoning for its active-parameter size.
  • 73.5% SWE-bench Verified, with Microsoft calling its SWE-bench Pro result toe-to-toe with a leading Anthropic model.
  • Preferred over Sonnet in blind human evaluations by Surge raters, per Microsoft.
  • Foundry list price undercuts both the Sonnet and Opus rates held in Hokai records.
Show 2 more
  • 97.0% on AIME 2025 and 94.5% on AIME 2026, strong math/scientific reasoning for a 35B-active model.
  • Preferred over Sonnet 4.6 in blind human evaluations across 1,276 tasks despite a smaller active-parameter footprint.

Limits

  • No vision, audio, or video input support.
  • Preview status: pricing and limits can change before general availability.
  • Terminal-Bench 2.0 (46.0%) and MultiChallenge (53.0%) trail top frontier agentic models.

MAI-Thinking-1 alternatives

The nearest AI models on Hokai for the same job.

ModelLabContextIn / out per 1MReads and writes
MAI-Thinking-1 This pageMicrosoft256K$2 / $8text → text
Hy4 flashTencent's Hy4 flash, released 16 November 2026, is a lower-cost multimodal model that reads text and images across a 1M-token window.Tencent1M$0.07 / $0.281text, image → text
Nemotron-Labs-SigmaNVIDIA's Nemotron-Labs-Sigma, released in November 2026 in 8B and 3B sizes, writes text by denoising whole blocks of tokens in parallel, not one at a time.NVIDIA4K--text → text
Qwen4 MaxAlibaba's Qwen4 Max, released 16 November 2026, is a sparse mixture-of-experts model with a 1M-token context and text, image and video input.Alibaba Cloud1M$1.50 / $4.50text, image, video → text

MAI-Thinking-1 specs

Context window
256K tokens
Architecture
Sparse Mixture of Experts (Latent MoE)
Mixture of experts
Yes
Parameters
35B active / ~1T total (base model: MAI-Base-1)

Training data

Trained from scratch on commercially licensed, clean, traceable data with no third-party model distillation. Staggered cutoffs by content type: web HTML to ~September 2025, web PDFs to December 2025, GitHub code to June 2025, books and journals to March 2026. Safety trained via the same reinforcement learning loop as capability, treating both unsafe compliance and unnecessary refusal as defects.

API endpoints

NameEndpointAuth
Microsoft Foundryhttps://microsoft.ai/models/mai-thinking-1/Azure AI Foundry API Key

SDKs

  • Python
  • TypeScript

Prompting patterns

Show 2 prompting patterns
  • Use developer system instructions for role framing

    Set a developer-role message defining task scope before the user turn.

    MAI-Thinking-1 supports explicit developer instructions separate from system/user turns, following the Chat Completions API convention.

  • Lean on function calling for structured output

    Define a JSON schema tool and force tool_choice for extraction tasks.

    Native function calling is a confirmed capability and produces more reliable structured output than free-text parsing.

Known quirks

Show 3 known quirks
  • No published max output token limit

    Microsoft has not disclosed a max output token ceiling for MAI-Thinking-1 alongside its 256K context window.

    Workaround: Test actual generation limits empirically in Foundry and cap output length explicitly in prompts or API parameters until Microsoft publishes an official figure.

  • Third-party price trackers disagree with the Azure list

    Some aggregators quote far lower per-token rates than the $2 input and $8 output on the Azure Foundry Models price list.

    Workaround: Budget from the Azure price list for the deployment type you will actually use, and re-check it before committing spend.

  • Staggered training data cutoffs by content type

    Web HTML stops around September 2025, PDFs December 2025, GitHub code June 2025, and books/journals March 2026, rather than one single cutoff date.

    Workaround: When freshness matters, ask the model to flag uncertainty on recent events and verify code-related knowledge against docs newer than June 2025.

MAI-Thinking-1 facts

Deployment
API · Azure
Licence
Proprietary (Microsoft Foundry Preview Terms)
Safety posture
Balanced
Released
02 Jun 2026
Current version
mai-thinking-1-20260602
Knowledge cutoff
01 Mar 2026
Family
MAI
Variant
Chat
Data retention
Not detailed on Microsoft's model page. Served through Microsoft Foundry, which offers enterprise governance and Azure data residency.

Content restrictions

  • Safety and refusal behavior trained jointly with capability via reinforcement learning, targeting both unsafe compliance and unnecessary over-refusal as defects.
  • Positioned under Microsoft's 'Humanist Superintelligence' framing: domain-bounded, not an unbounded autonomous entity.

Safety controls

Rlhf
Yes
Red teaming
not detailed on Microsoft's model page
System prompt
Yes
Constitutional AI
No
Content filter API
Yes

Data governance

Data residency options
azure regions via Foundry

Version history

  • mai-thinking-1-20260602 · 02 Jun 2026

    Initial private preview release on Microsoft Foundry. Base model MAI-Base-1. AIME 2025 97.0%, SWE-bench Verified 73.5%.

What changed

  • Release MAI-Thinking-1 launches in private preview on Microsoft Foundry (mai-thinking-1-20260602) source

Questions about MAI-Thinking-1

How much does MAI-Thinking-1 cost per 1M tokens?

Expect to pay $2 for input, $8 for output and $0.20 for cached input, each per million tokens, on the Global deployment while the model is in public preview (Azure list read October 23, 2026). Microsoft shows no free tier and no batch discount. A few aggregators quote far lower figures, but those do not match Azure's own table, so treat them as unverified. Rates can change before general availability.

MAI-Thinking-1 vs Claude Sonnet 4.6: which is the better pick?

On Microsoft's own figures MAI-Thinking-1 scores 73.5% on SWE-bench Verified and 84.2% on GPQA Diamond, and its Surge raters preferred it to Sonnet 4.6 in blind side-by-sides. On price it is cheaper: $2 and $8 per 1M tokens against $3 and $15 for Sonnet 4.6 in Hokai's records. Sonnet 4.6 offers a 1M-token window against 256K here. These benchmark numbers are vendor-reported and not independently reproduced, so run your own tasks before switching.

Is MAI-Thinking-1 open source or proprietary?

MAI-Thinking-1 is proprietary and closed-weight. There is no open-weights release and no self-hosting path. Access is through Microsoft Foundry, where it is in public preview, and Microsoft named Fireworks AI, Baseten and OpenRouter as third-party inference providers at launch. All of those routes serve the same hosted model, so commercial use follows Foundry's preview terms rather than a public license file.

Does MAI-Thinking-1 train on user data?

Microsoft's model page does not state a default training-on-inputs or retention policy for MAI-Thinking-1 specifically. It is served through Microsoft Foundry, which Microsoft markets with enterprise governance and Azure data residency. No model-specific SOC 2, ISO 27001 or HIPAA statement was found on the page. Teams in regulated industries should ask their Microsoft account team for written data-handling terms before sending sensitive data.

Who should pick MAI-Thinking-1, and who is better served elsewhere?

It suits enterprise teams already on Azure and Foundry who want a first-party reasoning model for coding, math and tool calling at $2 and $8 per 1M tokens. It is a weaker fit for vision or audio work, for long messy terminal automation (Terminal-Bench 2.0 is 46.0%), and for teams that need a general-availability SLA, since it is still in preview. Anyone needing a 1M-token window should look at GPT-5.5 or Gemini 3.1 Pro instead.

About MAI-Thinking-1

MAI-Thinking-1 is Microsoft AI's first flagship in-house reasoning model, announced at Microsoft Build on June 2, 2026. It is a sparse Mixture of Experts model with 35 billion active parameters out of roughly 1 trillion total, built on the internal MAI-Base-1 foundation model. Microsoft says it built the model from scratch on licensed data, without distilling any third-party model, which makes it the first serious attempt by Microsoft to ship frontier-class reasoning without leaning on a partner's weights.

Read the full background

The headline numbers are Microsoft's own, reported at launch: 97.0% on AIME 2025, 94.5% on AIME 2026, 84.2% on GPQA Diamond, 73.5% on SWE-bench Verified and 52.8% on SWE-bench Pro, which Microsoft describes as toe-to-toe with Claude Opus 4.6 on the Pro set. Terminal-Bench 2.0 (46.0%) and MultiChallenge (53.0%) are lower, so the model looks stronger at bounded math and coding tasks than at long, messy multi-turn terminal work. On the vendor page, Microsoft adds that independent raters on Surge preferred it to Claude Sonnet 4.6 in blind side-by-sides. None of these figures has been independently reproduced on a public leaderboard that Hokai tracks.

Pricing is now public: the Global deployment on Microsoft's Foundry price list is $2 for input and $8 for output per million tokens, read on October 23, 2026. Against Hokai's records that sits under Sonnet's $3 and $15, and well under Opus's $5 and $25. The model is a text-in, text-out reasoner with a 256,000-token context window, function calling and structured output; Microsoft has not published a maximum output limit. Foundry access has moved from private to public preview, and the price can still change before general availability.

Access is through Microsoft Foundry only, plus the third-party inference providers Microsoft named at launch (Fireworks AI, Baseten and OpenRouter). Baseten is the one Hokai lists as a service; for gateway routing, LiteLLM is a common layer. There are no open weights and no self-hosting path. Teams that want a cheaper Microsoft coding model should look at MAI-Code-1.1-Flash at $0.20 and $1.20.

Who should not pick it: anyone who needs vision, audio or video input, a 1M-token window (which GPT-5.5 offers, as do Google's Gemini 3.1 Pro and Anthropic's Opus 4.8), published compliance paperwork for this specific model, or a model that has left preview. The alternatives win on modality and context length; MAI-Thinking-1 wins when the buyer is already inside Azure governance and wants a first-party reasoning model at a mid-range price. If you are still choosing, Smart Match can shortlist models against your own budget and workload, and the model directory lists the rest.

Sources and guides

More AI Models on Hokai

  • Gemini 4 HydrogenGoogle DeepMind's November 2026 Gemini 4 model for coding agents, with four thinking levels from low to xhigh and a free tier in AI Studio.
  • Qwen4-Max-PreviewAlibaba's November 2026 Qwen 4 flagship preview, open only to Token Plan subscribers, with three reasoning levels from low to xhigh.
  • Claude Fable 5Superseded by Fable 5.1 and Fable 5.2 but still sold on the Claude API: Anthropic's first generally available Mythos-class model, released June 2026, with a 1M-token context.
  • Claude Fable 5.1Superseded by Claude Fable 5.2 in November 2026 but still sold on the Claude API: Anthropic's second Mythos-class model, released Sept 1, 2026, for long-horizon coding and research.
  • Gemini Decision 1Google DeepMind's decision model scores many questions over one multimodal input of up to 262,144 tokens and bills zero output tokens.
  • Shieldstral 1.0Mistral's open-weight multimodal safety classifier, released 4 August 2026, that screens text and images against policies you write.