Microsoft MAI Models and Where to Use Each One

Microsoft MAI models are the in-house family of artificial intelligence models developed by the Microsoft AI division for reasoning, programming, speech processing, and visual generation. Microsoft does not spell out what the MAI prefix stands for.

Microsoft released the models in waves through 2026, starting with seven at once on June 2.

You can reach most of them through Microsoft Foundry, the MAI Playground and GitHub Copilot. MAI-Code-1.1-Flash also runs on local hardware, and MAI-Cyber-1-Flash runs only inside the MDASH security harness.

The MAI Models at a Glance

Microsoft lists six MAI models as products: MAI-Thinking-1, MAI-Code-1.1-Flash, MAI-Image-2.6, MAI-Transcribe-2, MAI-Voice-2.1 and MAI-Cyber-1-Flash. The base models MAI-Base-2 and MAI-Base-1 are not offered as products.

Most models in the catalog can be tested in the MAI Playground and deployed through Microsoft Foundry, and MAI-Code-1.1-Flash runs in GitHub Copilot. Microsoft publishes dollar prices only for its speech-to-text and voice models. For the other product models it makes relative price claims and publishes no dollar figure on microsoft.ai.

• MAI-Thinking-1: Sparse Mixture of Experts (MoE) reasoning model for mathematics, programming, and enterprise work. Accessible in Microsoft Foundry and the MAI Playground. Microsoft calls its price mid-weight and publishes no dollar figure on microsoft.ai.
• MAI-Code-1.1-Flash: Small, agentic programming model that accepts screenshots, designs and mockups as input. In production in GitHub Copilot, with experimental access in Visual Studio Code (VS Code) due by the end of October 2026, and available as a local download. Token pricing is unpublished. Local calls carry no inference charges.
• MAI-Image-2.6 and MAI-Image-2.6-Flash: Image generation and editing models that work from text, images, web context and visual references. Accessible in Microsoft Foundry and the MAI Playground. Microsoft publishes no dollar figure on microsoft.ai, and Flash costs less than half the flagship's price.
• MAI-Transcribe-2 and MAI-Transcribe-2-Streaming: Speech-to-text recognition models covering 60 languages with diarization and code-switching capabilities. Accessible in Microsoft Foundry, MAI Playground, and OpenRouter. Published launch pricing is $0.10 per audio hour for batch processing and $0.54 per audio hour for streaming, both until the end of 2026.
• MAI-Voice-2.1 and MAI-Voice-2.1-Flash: Text-to-speech audio synthesis models offering voice cloning from a short reference clip and emotion control across 23 languages. Accessible in Microsoft Foundry, OpenRouter, MAI Playground, and Copilot Audio Expressions. Published rates are $22 per 1 million characters for standard synthesis and $15 per 1 million characters for Flash.
• MAI-Cyber-1-Flash: Compact security model that finds and fixes software vulnerabilities inside MDASH, Microsoft's multi-agent harness. Open only to verified defenders who register interest. No dollar price is published, and Microsoft claims comparable performance at 50 percent of the cost of leading models.
• MAI-Base-2: Foundation base model with 2.8 trillion total parameters natively pre-trained with vision. Access is not yet available to the public, with Microsoft planning partner weight licensing through Microsoft Foundry. Commercial pricing and context windows are currently unpublished.
• MAI-Base-1: The previous base model, which alternated dense and MoE layers and is the model MAI-Thinking-1 was built on. Microsoft calls it a strong model for its size and does not offer it as a product.

Origin and Structure of the In-House Lineup

Microsoft does not spell out the MAI acronym in its official posts, using the prefix for models built in-house by Microsoft AI. A June 2, 2026 launch post bylined by Mustafa Suleyman introduced the family through seven initial models.

That original June 2026 release included MAI-Thinking-1, MAI-Code-1-Flash, MAI-Image-2.5, MAI-Image-2.5-Flash, MAI-Transcribe-1.5, MAI-Voice-2, and MAI-Voice-2-Flash, as detailed in the seven-model launch announcement. Newer versions have since replaced several of them: MAI-Image-2.6, MAI-Code-1.1-Flash, MAI-Transcribe-2 and MAI-Voice-2.1. Microsoft also offers custom fine-tuning through Microsoft Frontier Tuning. The Microsoft AI models page says Microsoft Frontier Tuning fits a model to your business with full privacy and control.

Microsoft's June post names Microsoft Foundry, OpenRouter, Fireworks and Baseten as places to get the models. Microsoft invites people to try them in the MAI Playground. No MAI model is open source as of October 2026, and Microsoft plans to license MAI-Base-2 weights to trusted partners through Microsoft Foundry.

MAI-Thinking-1 Reasoning Architecture and Benchmarks

MAI-Thinking-1 is Microsoft AI's first reasoning model, a general-purpose model aimed at math, coding and enterprise work. MAI-Thinking-1 is a sparse MoE model with roughly 1 trillion total parameters with 35 billion active parameters per forward pass. A context window of 256,000 (256K) tokens accommodates extensive source material, and the model works with the Chat Completions application programming interface (API).

First announced during the seven-model launch in June 2026, MAI-Thinking-1 entered public preview inside Microsoft Foundry on August 12, 2026, as documented in Microsoft's MAI-Thinking-1 announcement. Microsoft says MAI-Thinking-1 used no distillation from third-party models. In an October 15, 2026 post, Microsoft said that MAI-Thinking-1 was built on MAI-Base-1.

Microsoft's published evaluations report an accuracy score of 97.0 percent on the AIME 2025 benchmark and 94.5 percent on AIME 2026. On SWE-Bench Pro, Microsoft says MAI-Thinking-1 goes toe-to-toe with Claude Opus 4.6. Microsoft gives no score for that test. Microsoft also says independent raters on Surge preferred MAI-Thinking-1 to Claude Sonnet 4.6 for overall quality in blind side-by-side tests.

MAI-Code-1.1-Flash Specifications and Local Deployment

MAI-Code-1.1-Flash is a specialized, compact agentic programming model built specifically for integration inside GitHub Copilot and Visual Studio Code (VS Code). Microsoft announced MAI-Code-1.1-Flash on October 7, 2026. It replaces MAI-Code-1-Flash, which launched on June 2, 2026 at Microsoft Build, per the June MAI-Code-1-Flash announcement. MAI-Code-1.1-Flash accepts screenshots, designs and mockups as input. Its languages include Python, C++, CSS, HTML, .NET, Java, JavaScript and TypeScript.

Performance metrics published in Microsoft's MAI-Code-1.1-Flash release announcement compare MAI-Code-1.1-Flash with the June 1.0 release. Microsoft states that MAI-Code-1.1-Flash costs one-quarter the price of the original model and uses 25 percent fewer tokens per coding task. MAI-Code-1.1-Flash also streams output tokens 25 percent faster within GitHub Copilot. Microsoft also reports a 22 percent improvement on Terminal-Bench 2.1 in the GitHub Copilot Command-Line Interface (CLI) and a 15 percent performance lift on specialized .NET development tasks.

For the earlier 1.0 model, Microsoft reported a SWE-Bench Pro score of 51.2 percent, compared to 35.2 percent for Claude Haiku 4.5. In addition to cloud deployment inside production GitHub Copilot, Microsoft also lets developers download MAI-Code-1.1-Flash and run it locally. Microsoft recommends a device with more than 120 gigabytes (GB) of random-access memory (RAM). The local version uses 3-bit quantization and has a 256K context window. Local calls carry no inference charges.

MAI-Image-2.6 Visual Generation and Image Editing

MAI-Image-2.6 creates and edits images from text, images, web context and visual references. Announced on September 4, 2026, the lineup consists of the flagship MAI-Image-2.6 model alongside a lower-cost MAI-Image-2.6-Flash model. The release updates the preceding MAI-Image-2.5 and MAI-Image-2.5-Flash models originally introduced in the June 2026 launch.

Microsoft pitches the flagship for quality generation and precise editing. MAI-Image-2.6-Flash gives production-ready quality at less than half the flagship's price. Microsoft's MAI-Image-2.6 announcement says the flagship held the number two position on the Arena leaderboard for both text-to-image synthesis and image editing as of September 4, 2026. On the Artificial Analysis platform, Microsoft reported a number two ranking for text-to-image generation and the number one ranking for image editing.

Developers can evaluate both visual models within the MAI Playground or deploy them for automated pipelines through public preview access in Microsoft Foundry. Microsoft has not published explicit dollar rates per image generation on microsoft.ai.

MAI-Transcribe-2 Speech Recognition and Streaming Performance

MAI-Transcribe-2 is a multilingual speech-to-text recognition model supporting 60 languages with speaker diarization, timestamps and domain biasing. Unveiled on September 3, 2026, the model identifies the spoken language automatically and handles code switching, such as Hinglish and Spanglish.

On technical evaluation leaderboards, Microsoft reports that MAI-Transcribe-2 took first place on the FLEURS benchmark across 60 languages with an average Word Error Rate (WER) of 5.2 percent. The model also secured the number two position on the Artificial Analysis WER leaderboard, as documented in the MAI-Transcribe-2 announcement. Microsoft established an introductory launch price of $0.10 per audio hour valid through the end of 2026.

For live audio, Microsoft introduced MAI-Transcribe-2-Streaming on October 1, 2026, delivering live transcription in 60 languages at $0.54 per audio hour through the end of 2026. In Microsoft's streaming transcription announcement, Microsoft stated MAI-Transcribe-2-Streaming debuted at number one for transcription accuracy across both partial and final outputs on Artificial Analysis. Both transcription variants are accessible through Microsoft Foundry, the MAI Playground, and OpenRouter.

MAI-Voice-2.1 Audio Synthesis and Latency Tiers

MAI-Voice-2.1 and MAI-Voice-2.1-Flash provide automated text-to-speech generation across 23 languages and 26 distinct geographic locales. Announced on October 1, 2026, as successors to the June 2026 MAI-Voice-2 release, the models allow voice cloning from a short reference audio sample without requiring downstream fine-tuning. Both models also offer emotion control.

The two voice models differ in inference latency. Official specifications on the Microsoft AI models page list an inference latency of approximately 550 milliseconds (ms) for MAI-Voice-2.1, which Microsoft points to for audiobooks, content creation and voice-over. Conversely, MAI-Voice-2.1-Flash has an inference latency of about 45ms, making it suitable for real-time contact center voice agents, interactive voice response (IVR) telephony, and interactive virtual assistants.

Microsoft charges $22 per 1 million characters for MAI-Voice-2.1 and $15 per 1 million characters for MAI-Voice-2.1-Flash. Both models are available through Microsoft Foundry under Azure Speech, OpenRouter, and the MAI Playground. In consumer software, Microsoft puts the voice models into Copilot through a feature named Copilot Audio Expressions.

MAI-Cyber-1-Flash Automated Vulnerability Remediation

MAI-Cyber-1-Flash is a compact, code-intensive cybersecurity model derived from the MAI-Thinking-1 architectural lineage to identify and resolve software vulnerabilities. Announced on August 13, 2026, the model runs inside the MDASH harness. MDASH coordinates more than 100 agents to find, validate and prove vulnerabilities. MAI-Cyber-1-Flash handles up to 90 percent of MDASH tasks.

Microsoft's MAI-Cyber-1-Flash announcement reports results on the CyberGym security benchmark. When paired with GPT-5.4 inside MDASH, the system achieved a 96 percent score on the any-crash evaluation metric and an 86.3 percent score on final submission. Microsoft says MAI-Cyber-1-Flash performs comparably to leading models at 50 percent of their cost.

Only verified defenders can use MAI-Cyber-1-Flash, through MDASH, after registering interest with Microsoft. Microsoft names no API or Foundry listing for the model. Pricing details remain unlisted beyond Microsoft's comparative cost claims.

MAI-Base-2 and MAI-Base-1 Foundation Model Architectures

MAI-Base-2 is Microsoft AI's in-house base model. On October 15, 2026, Microsoft announced that its pre-training is finished and post-training is underway. MAI-Base-2 is an MoE model with 2.8 trillion total parameters and 80 billion active parameters. The model was natively pre-trained with vision on 50 trillion tokens and has a March 2026 knowledge cutoff.

Microsoft reports that MAI-Base-2 outperforms every open-weight baseline evaluated on held-out tests and exhibits 4 to 12 times greater pre-training floating-point operations (FLOPs) efficiency than DeepSeek-V4-Pro-Base. More detail is in the MAI-Base-2 explained guide and the MAI-Base-2 vs DeepSeek V4 Pro comparison.

MAI-Base-2 is not yet accessible for commercial deployment, offering no public API, application download, or confirmed release date. Microsoft plans to license the weights to trusted partners through Microsoft Foundry. Preparation for MAI-Base-3 is underway. The earlier MAI-Base-1 combined dense and MoE layers to provide the underlying foundation for MAI-Thinking-1, though Microsoft never released MAI-Base-1 as an independent commercial product.

How Much Microsoft MAI Models Cost

Only four Microsoft MAI models have a published dollar price as of October 2026. For speech-to-text transcription, MAI-Transcribe-2 carries an introductory rate of $0.10 per audio hour, while MAI-Transcribe-2-Streaming costs $0.54 per audio hour. Both launch prices run through the end of 2026. For text-to-speech audio production, MAI-Voice-2.1 charges $22 per 1 million characters, compared to $15 per 1 million characters for MAI-Voice-2.1-Flash.

For MAI-Thinking-1, MAI-Code-1.1-Flash, MAI-Image-2.6 and MAI-Cyber-1-Flash, Microsoft has not published dollar prices on the Microsoft AI models page. Microsoft calls MAI-Thinking-1's price mid-weight. MAI-Code-1.1-Flash costs a quarter of the June 2026 model's price and uses 25 percent fewer tokens per task. MAI-Image-2.6-Flash is priced at less than half the cost of the flagship image model. Microsoft says MAI-Cyber-1-Flash performs comparably to leading models at 50 percent of their cost.

Running MAI-Code-1.1-Flash on your own hardware carries no inference charges, and Microsoft recommends more than 120GB of RAM for it. Prices for other AI models are on the AI model pricing page.

How to Try or Deploy Microsoft MAI Models

Where you can use a Microsoft MAI model depends on the model: Microsoft Foundry, MAI Playground, GitHub Copilot, OpenRouter or MDASH. You can try MAI-Thinking-1, MAI-Image-2.6, MAI-Transcribe-2, and MAI-Voice-2.1 inside the MAI Playground. Microsoft Foundry hosts most of the models, several of them in public preview. Microsoft also offers custom fine-tuning through Microsoft Frontier Tuning.

MAI-Code-1.1-Flash is in production in GitHub Copilot. Experimental access in the GitHub Copilot app, GitHub Copilot CLI and VS Code is due by the end of October 2026. OpenRouter carries MAI-Transcribe-2 and MAI-Voice-2.1, and Microsoft's June launch post also names Fireworks and Baseten as places to get MAI models. You can also download MAI-Code-1.1-Flash and run it locally, with no inference charges for local calls.

MAI-Cyber-1-Flash runs only inside MDASH. Verified defenders must register interest with Microsoft to use it. Microsoft plans to license MAI-Base-2 weights to trusted partners through Microsoft Foundry.

Workload Selection and Model Routing Across the Family

Pick an MAI model by the type of input and output your job needs. For math, coding and enterprise reasoning tasks, use MAI-Thinking-1 through the Chat Completions API. For coding inside GitHub Copilot, use MAI-Code-1.1-Flash. MAI-Code-1.1-Flash reads screenshots and mockups, and Microsoft says it streams tokens 25 percent faster in GitHub Copilot than the June model.

For speech to text, use MAI-Transcribe-2 for recorded audio and MAI-Transcribe-2-Streaming for live transcripts. For audiobooks, content creation and voice-over, Microsoft points to MAI-Voice-2.1. For call-center agents, voice assistants and IVR, Microsoft points to MAI-Voice-2.1-Flash, with about 45ms of model latency.

Other speech models are compared in the best speech-to-text APIs review and the best text-to-speech for voice agents comparison. Microsoft's wider lineup is covered in the Microsoft AI tools for business guide, and Copilot plan costs are in the GitHub Copilot pricing breakdown.

Technical Milestones and Future Product Roadmap

Several changes are coming for Microsoft MAI models. Microsoft is post-training MAI-Base-2, plans to license its weights to trusted partners through Microsoft Foundry and is preparing MAI-Base-3.

Microsoft aims to complete experimental rollout of MAI-Code-1.1-Flash across the GitHub Copilot application, the Copilot CLI, and VS Code by the end of October 2026. The launch prices for MAI-Transcribe-2 ($0.10 per hour) and MAI-Transcribe-2-Streaming ($0.54 per hour) run only until the end of 2026.

Before you build on one of the Microsoft MAI models, check its current status on the official Microsoft AI models page.