Microsoft Unveils Seven AI Models at Ignite and Says Its Coding Model Hits 98% of GPT-6.1 Sol’s Quality at 40% of the Price

Microsoft said MAI Code 2 reaches 98% of GPT-6.1 Sol’s coding quality at 40% of the price, a comparison that uses a promotional rate, at Ignite Tuesday.

Share this article

Featured image: Building 92 on Microsoft’s Redmond, Washington, campus, in a June 2015 file photo. The photo shows the Redmond campus, not the San Francisco Ignite conference. Photo: Jiaqian AirplaneFan, Wikimedia Commons, CC BY 3.0. Cropped to 16:9.

Microsoft on Tuesday, Nov. 17, 2026, released a family of seven new in-house AI models and said MAI Code 2, its main coding model, reaches 98% of the quality of OpenAI’s GPT-6.1 Sol at about 40% of the price. Those coding-price figures use a 50% promotional rate, Microsoft disclosed in footnotes to its own announcement.

Advertisement

The models were unveiled at Microsoft Ignite in San Francisco, Reuters reported. The same day, Chairman and CEO Satya Nadella announced an AI Transparency Initiative with more than 40 organizations, including Accenture, Adobe, KPMG, Mastercard and Volvo Cars, aimed at a common approach to documenting what AI agents do inside companies.

Models announced
7
Plus the Maia 200 accelerator, Microsoft said
MAI Code 2 vs. GPT-6.1 Sol
98% / 40%
Quality / price, Microsoft said, at a promotional rate
Transparency partners
40+
Organizations in the new initiative, Microsoft said
Dragon Copilot use
10 million
Patient encounters a month, Microsoft said

What Microsoft says the seven models can do

In a Microsoft AI post dated Nov. 17, the company listed seven models:

  • MAI Cyber 2 Flash, which it said matches GPT-6 Astra’s quality at less than 2% of the cost per task on the CyberGymE2E benchmark.
  • MAI Code 2, which it said reaches 98% of GPT-6.1 Sol’s coding quality at about 40% of the price, and 96% of Claude Opus 5.5’s performance at 5% of the price.
  • MAI Code 2 Flash, which it said offers comparable performance to GPT-6 Luna at 65% of the price and to Claude Haiku 5.5 at 20% of the price.
  • MAI Image 2.7, which it said ranks first on Artificial Analysis image editing and outperforms GPT Image 2.6 Sunburst (max) at one fifth of the price.
  • MAI Realtime 1, a speech-to-speech model that it said outperformed GPT-Realtime 2.1 in its evaluations and starts at $0.01 per audio minute.
  • Dragon Copilot Transcribe, built on MAI Transcribe 2, which it said already powers 10 million patient encounters a month, halves error rates in multilingual workflows, and is used by a third of U.S. physicians.
  • Decision 1.1, a routing model that it said is used by 4,000 customers and is 50 times faster and 79 times cheaper than GPT-Sol 6.1.

Microsoft also said its second-generation accelerator, Maia 200, is now in production and delivers 1.7 times more tokens per watt. A companion post on MAI Code 2 said the coding models are available in GitHub Copilot, and that Code 2 Flash is in limited-access preview in Microsoft Foundry.

The company said the coding models are “explicitly not yet tuned for harder problems, like fundamental science research and very long horizon agentic work.” In a blind frontend evaluation, it said, reviewers preferred MAI Code 2 over GPT-6 Sol 64% of the time, in a test it said was annotated by 258 people.

Reuters, interviewing Microsoft AI chief Mustafa Suleyman, quoted him as saying, “This isn’t just one model. It’s not a flash in the pan. This is a system, a hill-climbing machine that we have built, which co-optimizes performance and price.” Suleyman told Reuters that building models for specific domains was helping the company keep costs low while matching rivals, and that alignment — preventing harmful actions and keeping models on their intended behavior — was his top priority. He pointed to the Hugging Face incident involving OpenAI agents as part of the backdrop for that focus, Reuters reported.

What the data shows

The headline coding-price claims are Microsoft’s, not an independent audit by The Boston Report. In footnotes to the MAI Code 2 post, Microsoft said the Code 2 and Code 2 Flash cost comparisons were run at maximum reasoning effort on four or three named coding benchmarks, and that “MAI-Code-2 cost is shown at its 50% promotional price,” with the same note for Code 2 Flash. FrontierCode 1.1 scores and costs, it said, came from Artificial Analysis.

A Boston Report calculation from that footnote is straightforward: if the published Microsoft price in each ratio is half of the non-promotional price, the share of the rival’s price would be twice the stated figure. On that reading, MAI Code 2 would be about 80% of GPT-6.1 Sol’s price and about 10% of Claude Opus 5.5’s price, rather than 40% and 5%. MAI Code 2 Flash would be about 130% of GPT-6 Luna’s price — higher than the rival, not lower — and about 40% of Claude Haiku 5.5’s price, rather than 65% and 20%. Those doubled figures are a TBR calculation from Microsoft’s footnote. They assume the 50% promotion applies only to Microsoft’s side of the ratio and that the rival prices used in the comparison are not also discounted. Microsoft did not publish a table of non-promotional list prices alongside the promotional comparisons.

Grouped bar chart of Microsoft-stated MAI Code 2 and Code 2 Flash prices as a share of named rivals, next to Boston Report calculations that double those shares to show an implied list price without the 50% promotion
Source: Microsoft AI, “MAI Code 2 and Code 2 Flash,” Nov. 17, 2026, footnotes 1 and 3. Implied list shares are Boston Report calculations (stated share × 2). Chart: The Boston Report.

Other figures in the Ignite post were not accompanied by the same promotional-price footnote. Those include the claim that MAI Cyber 2 Flash costs less than 2% as much as GPT-6 Astra per task, that MAI Image 2.7 is priced at one fifth of GPT Image 2.6 Sunburst (max), and that Decision 1.1 is 79 times cheaper than GPT-Sol 6.1. Microsoft also said output-token prices for MAI Cyber 2 Flash are 86% lower than for MAI Cyber 1.1 Flash, and that MAI Cyber accounts for almost 40% of scans in MDASH, with 90% of scans that started on a MAI model continuing to use one. Those are company figures.

A new transparency initiative for AI agents

Alongside the model release, Microsoft said it is launching an AI Transparency Initiative because organizations can often see activity around a model but cannot reconstruct what happened between a request and an outcome. “Greater AI autonomy requires greater transparency,” the company said. Hayete Gallot, executive vice president of Microsoft Security, said, “For security controls to be effective, every organization needs greater transparency into the model itself.”

The initiative has three parts, Microsoft said. First, it will expand its work with the National Institute of Standards and Technology’s SI Agent Standards Initiative. Second, it is working with more than 40 organizations to develop open standards so AI systems can be independently verified. Third, it will build the resulting evidence into Microsoft AI, Microsoft Foundry and Microsoft Copilot, so customers can use it with their own security and compliance tools.

Advertisement

Aanchal Gupta, Adobe’s chief security officer, said security teams “need clear visibility into what those agents are accessing, what actions they’re taking and how those actions can be verified.” John Israel, global chief information security officer at KPMG International, said clients “shouldn’t have to choose between innovation and safety.” Harpreet Sidhu, global cybersecurity lead at Accenture, called the effort an “open, industry-driven approach to AI transparency.”

Microsoft framed the work as “see, verify, control”: know what an agent was asked to do and what it did; do not treat the model as the sole source of truth about its own behavior; and keep control over what data and actions are permitted, including when the underlying model changes. Accountability, the company said, “stays with the organization.”

The emphasis on watching what agents actually do sits in a week of related reporting. Google said it had found an AI framework that writes and runs exploits from software patches, and Apple moved to tighten Mac permissions for AI agents in macOS 27.3. Microsoft’s initiative is a standards and product effort, not a regulator’s rule, and the company did not announce a date by which the new evidence tools will be in customers’ hands.

What is still company-reported

The performance and price comparisons above are Microsoft’s, from its own benchmarks and, in the coding cases, from Artificial Analysis figures it cited. Reuters described the new suite as part of a race with Anthropic and OpenAI and said Microsoft also touted a cyber-specific model that it plans to offer to a limited group of customers. Microsoft’s own posts said MAI Code 2 is in GitHub Copilot now and that Code 2 Flash is in a limited-access Foundry preview; they did not use the phrase “limited group” for MAI Cyber 2 Flash.

Microsoft also announced MAI Licensed Weights, which it said will let customers customize a frontier MAI model and run it in their own Azure environment, and it cited an AT&T custom model for 5G network analytics that, it said, outperformed GPT-6.1 Sol at 50 times lower cost. Those claims were not accompanied by an independent evaluation in the announcement.

Sources

Read next in Artificial Intelligence

Get The Boston Report Briefing

Our free weekly email with the most important stories from Boston and the world. Unsubscribe anytime.

Free. We never sell your address. Privacy policy

Get The Boston Report Briefing
Our free weekly email with the most important stories from Boston and the world. Unsubscribe anytime.
Sign up