Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On Radar3DNews AI→ original

Microsoft Releases MAI Code 2 Coding Model: 98% of GPT-6.1 Sol's Quality at 40% of the Price

At Ignite 2026, Microsoft showed MAI Code 2, its model for writing and editing code, which is positioned as a cheaper replacement for rivals' expensive models. According to the company's tests, it delivers 98% of GPT-6.1 Sol's performance at roughly 40% of the cost, and 96% of Claude Opus 5.5's performance for 5% of the cost. It is already available in GitHub Copilot. These figures come from Microsoft itself, so test the model on your own code before relying on them.

AI-processed from 3DNews AI; edited by Hamidun News
Microsoft Releases MAI Code 2 Coding Model: 98% of GPT-6.1 Sol's Quality at 40% of the Price
Source: 3DNews AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary source3DNews AI ↗
Editor-in-Chief
Key metric #1In a blind interface test, developers chose its output over GPT-6 Sol's in 68% of cases
Key metric #2MAI Code 2 Flash, for small edits, costs about 20% of Claude Haiku 5.5
Key metric #3MAI Cyber 2 Flash finds vulnerabilities for less than 2% of the competitor's cost

Paying for an AI coding assistant is expensive, especially when agents run hundreds of tasks against repositories every day. At its Ignite 2026 conference, Microsoft offered an alternative: its own MAI Code 2 model for writing and fixing code. According to the company's tests, it delivers 98% of the performance of GPT-6.1 Sol (OpenAI's model) at about 40% of its price. Put simply, if GPT-6.1 Sol scores "100 points" on your set of tasks and costs 100 rubles, MAI Code 2 scores about 98 points for 40 rubles. This is only the most prominent of seven new models in the MAI family.

What MAI Code 2 does and what it costs

Against Anthropic's Claude Opus 5.5, the gap is even wider: 96% of the performance at just 5% of the competitor's cost. The testing covered typical enterprise tasks, code work, and software tooling.

Separately, Microsoft ran a blind comparison on interface creation. Developers did not know which model produced which output, and in 68% of cases they preferred MAI Code 2's result over GPT-6 Sol's. The tasks involved buttons, control panels, and visualizations, the front-end routine that is most often handed to AI.

For high-volume small operations there is MAI Code 2 Flash: explaining a piece of code, fixing a bug, and quickly extending a project. According to the company, it is comparable to GPT-6 Luna at 65% of the cost and to Claude Haiku 5.5 at about 20% of the cost. MAI Code 2 is available through GitHub Copilot, while the Flash version is in limited preview in Microsoft Foundry.

Microsoft itself named the limitation: the model is tuned for standard programming, not for fundamental scientific research or long, autonomous AI agent work.

The rest of the lineup: security, images, voice

MAI Cyber 2 Flash is built for vulnerability discovery. On the CyberGymE2E benchmark, Microsoft says its quality is comparable to GPT-6 Astra, and running the task costs less than 2% of the competitor's cost. Compared with the previous MAI Cyber 1.1 Flash, the output token price fell by 86%. Earlier Cyber versions already account for nearly 40% of checks in the MDASH system, and 90% of customers who started scanning with them kept using them.

MAI Image 2.7 took first place in the Artificial Analysis image editing ranking, beating GPT Image 2.6 Sunburst in its maximum configuration at roughly one-fifth the price. In text-to-image generation it ranks fourth overall, which Microsoft counts as second place among AI developers. The Flash version is built for volume: about 1,000 images for around $10.

For voice, Microsoft released MAI Realtime 1: according to internal and external tests, including Dynamics 365, it outperformed GPT-Realtime 2.1, and processing costs from $0.01 per minute of audio. MAI Transcribe 2 underpins Dragon Copilot Transcribe, which documents conversations between doctors and patients: about 10 million encounters per month, with half as many errors in multilingual transcription. The service is used by roughly a third of American doctors.

Router, own weights, and own chip

Decision 1.1 is a router that selects a model for each request: simple requests go to a fast, cheap model, complex ones to an expensive one. Microsoft claims it runs 50 times faster and costs 79 times less than GPT-Sol 6.1. About 4,000 customers use it, and on the Artificial Analysis Decision Index it ranks ahead of solutions from Google and OpenAI.

The MAI Licensed Weights program lets companies obtain Microsoft's base models, fine-tune them on their own data, and run them in secure Azure infrastructure. The customer controls access, and the data and fine-tunes stay with them. Underpinning all of this is the Maia 200 accelerator: it is already in production and delivers 1.7 times more processed tokens per watt than Maia 100.

What to do in practice

The key caveat: almost all of the figures come from Microsoft itself, and there are no independent checks yet. But the direction is clear. If your team pays for Copilot or for its own agents, assemble 30–50 typical tasks from a real repository and run MAI Code 2 alongside your current model. If the quality difference turns out to be within a few percent, the 2.5–20x savings justify the switch, at least for routine edits. Complex, long autonomous scenarios are best left to heavier models for now, as Microsoft itself candidly warns.

Story

Key Questions Answered

What MAI Code 2 does and what it costs?

Against Anthropic's Claude Opus 5.5, the gap is even wider: 96% of the performance at just 5% of the competitor's cost. The testing covered typical enterprise tasks, code work, and software tooling. Separately, Microsoft ran a blind comparison on interface creation. Developers did not know which model produced which output, and in 68% of cases they preferred MAI Code 2's result over GPT-6 Sol's. The tasks involved buttons, control panels, and visualizations, the front-end routine that is most often handed to AI.

What are the key details on The rest of the lineup: security, images, voice?

MAI Cyber 2 Flash is built for vulnerability discovery. On the CyberGymE2E benchmark, Microsoft says its quality is comparable to GPT-6 Astra, and running the task costs less than 2% of the competitor's cost. Compared with the previous MAI Cyber 1.1 Flash, the output token price fell by 86%. Earlier Cyber versions already account for nearly 40% of checks in the MDASH system, and 90% of customers who started scanning with them kept using them.

What are the key details on Router, own weights, and own chip?

Decision 1.1 is a router that selects a model for each request: simple requests go to a fast, cheap model, complex ones to an expensive one. Microsoft claims it runs 50 times faster and costs 79 times less than GPT-Sol 6.1. About 4,000 customers use it, and on the Artificial Analysis Decision Index it ranks ahead of solutions from Google and OpenAI. The MAI Licensed Weights program lets companies obtain Microsoft's base models, fine-tune them on their own data, and run them in secure Azure infrastructure.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…