The Specialty News
AI

Microsoft Unveils Seven Homegrown AI Models, Coding at 5% of Opus's Price

MAI Code 2 scores 96% of Claude Opus 5.5 on four coding benchmarks and ships today inside GitHub Copilot.

By The Specialty News DeskEdited by 4 min read
Microsoft Unveils Seven Homegrown AI Models, Coding at 5% of Opus's Price
Photo: Microsoft

Microsoft's new coding model scores 96% of what Claude Opus 5.5 does, at 5% of the price. MAI Code 2 was one of seven models Microsoft AI unveiled on Tuesday at Ignite in San Francisco, all built in-house. For developers it is available today in GitHub Copilot, and it turns near-frontier coding into something a team can leave running all day.

Near-Frontier Code for a Twentieth of the Bill

Microsoft measured MAI Code 2 on four public coding tests, DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, and averaged the results. Against OpenAI's GPT 6.1 Sol it reaches 98% of the score at about 40% of the price. Against Anthropic's Claude Opus 5.5 it reaches 96%, for a twentieth of the cost.

The price comes with a footnote worth reading: Microsoft lists MAI Code 2 at a 50% promotional rate. Even at full price, the gap stays wide enough to change what a team can afford to automate.

Its smaller sibling, MAI Code 2 Flash, is built for volume. Microsoft puts it level with GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of their prices. Both models are live in GitHub Copilot today, and Flash is also in a limited preview on Microsoft Foundry.

Code was only the opening act. The other models on stage went after hospitals, design studios and call centres.

A Model for Every Job, From Clinics to Call Centres

MAI Image 2.7 ranks first in Artificial Analysis' image-editing leaderboard, with 1,215 Elo points from 5,478 votes, four ahead of OpenAI's GPT Image 2.6. It costs about $35 per 1,000 images, and a Flash version about $10, both in public preview on Foundry and through fal, OpenRouter, Vercel and Genspark.

MAI Realtime 1 handles live speech-to-speech conversations starting at $0.01 per audio minute, and it speaks the same API as OpenAI's Realtime, so developers can switch without rewriting their apps. MAI Cyber 2 Flash, with a one-million-token context window, arrives in Microsoft's security products from mid-December. Decision 1.1, a model that routes each request to the right tool, is already used by 4,000 customers.

The quietest number came from medicine. Dragon Copilot Transcribe, built on MAI Transcribe 2, now writes up doctors' visits at hospital scale and halves the error rate in multilingual conversations.

That leaves the most unusual offer of the day: Microsoft will hand over the model itself.

Your Weights, on Your Own Azure

With MAI Licensed Weights, a company can take a frontier MAI model, tune it with its own data and run it inside its own Azure environment. For a bank or a hospital that cannot send records to someone else's servers, that is the difference between watching AI from the sidelines and putting it to work.

Underneath sits Microsoft's own chip. Maia 200 is now in production running the company's largest models, with 1.7 times more tokens per watt. Owning the model and the silicon is what lets Microsoft quote prices like these.

Satya Nadella framed the day around that idea in his keynote.

“Intelligence, by definition, needs to be broadly distributed versus just collapse into a couple of models.”— Satya Nadella

For developers, the practical change starts today: the coding agent in their editor now has a Microsoft model behind it that costs a fraction of the frontier. When near-frontier code costs a twentieth as much, the question stops being whether to automate a task and becomes which one to start with.

Near-Frontier Code, Fraction of Cost

A visual summary of this story

The Brief

Stay curious

AI and technology: what changes and why it matters.
Your daily selection, in English or Spanish.

Free forever. Unsubscribe anytime.

Conversation

Start the conversation

No account needed. Comments are checked automatically — keep it civil.

More stories

Keep reading