96% of the performance of Claude Opus 5.5 for 5% of its price. Microsoft AI highlighted this ratio to present MAI Code 2, its new programming model, on November 17 at Ignite, its conference for businesses, in San Francisco. The figure comes from Microsoft. It corresponds to an average of four code benchmarks, DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, which according to the publisher target common tasks observed in GitHub Copilot. The Microsoft AI page does not detail the protocol.
Microsoft alone calculates the gap with Claude and GPT
The same text places MAI Code 2 at 98% of the quality of GPT-6.1 Sol, from OpenAI, for around 40% of the price. The two ratios do not correlate well with public prices. The grid of GitHub, a subsidiary of Microsoft, charges Claude Opus 5.5 to 20 dollars per million tokens produced and GPT-6.1 Sol at 10 dollars. If the percentages were based on the unit price, a model sold for 40% of the price of GPT-6.1 Sol would cost approximately 20% of the price of Opus 5.5. Microsoft’s “price” therefore probably refers to the total cost of running the tests, which also depends on the number of tokens each model consumes, but the page does not say this, and MAI Code 2 is not included in the GitHub grid. ActuIA showed on November 17, with Claude Fable 5.2, how much the price per task can deviate from the price per token. As long as this rate is not published, the 5% ratio remains a commercial argument that no one can remake.
MAI Code 2 is available in GitHub Copilot. Its faster and cheaper version, MAI Code 2 Flash, is also fast, and opens with limited access in Microsoft Foundry, Azure’s template platform. The other new features target images, with MAI Image 2.7 Flash at around $10 for 1000 images, and voice, with MAI Realtime 1, starting at $0.01 per minute, testable in MAI Labs. Ignite’s Book of News ranks MAY Image 2.7 in preview.
An offer targets major accounts. With MAI Licensed Weights, “trusted customers” will be able to obtain the weights of a MAI model and run it in their own Azure environment, says the official Microsoft blog, without price. The same text announces the arrival of GPT 6.2, from OpenAI, in Foundry. At the beginning of June, the Build conference had already been used to launch seven MAI models.
Autopilot receives its own identity in the company directory
Presented at the end of September in early access, Copilot Autopilot begins to be deployed to customers of Frontier, the early access program of Microsoft 365. The agent is attached to a user and runs permanently in the Microsoft 365 environment of the company, with its memory and its tools. It monitors Teams channels, relaunches emails and executes recurring tasks without waiting to be requested. Microsoft does not give a general availability date or price.
Governance is the most detailed part. Each Autopilot receives a named manager and their own identifier in Microsoft Entra, the directory where the company already manages the accounts of its employees, details Jared Spataro on the Copilot blog. The IT department can restrict or block access at any time, and its connections and actions are logged under its own name. In the audit logs, what the agent does is distinguished from what the employee he assists does. This agent identity is only in preview for Microsoft’s cloud agents, according to the Book of News, as well as for agents installed on workstations, starting with GitHub Copilot CLI and Claude Code from Anthropic.
Microsoft describes a pricing model without numbers. Jared Spataro mentions a subscription per user, supplemented by usage-based billing for the heaviest tasks. Agent 365, the agent management console sold for $15 per user per month since May 1, according to Redmond Magazine, receives spending caps in preview, and cost forecasts opened to Frontier customers in November and more widely in December. Copilot Code, which builds apps based on a plain language description, is scheduled to go into general availability in December.
Azure customers can lease Cobalt 200 now
Virtual machines based on the Cobalt 200 processor have been available since November 17. The chip aligns 132 Arm Neoverse V3 cores engraved in 3 nm by TSMC, and Microsoft measures up to 50% more performance than on Cobalt 100. Amadeus says it has gained more than 30% in throughput on flight searches. The 14 launch regions include four European sites, in Italy, Spain, Sweden and the United Kingdom, none in France. The Azure blog promises 23 by the end of 2026, the Book of News 24. The announcement does not display any prices and refers to the general grid of virtual machines.
Maia 300 follows a different schedule. Unveiled on October 12, the N2P accelerator at TSMC features 288 GB of HBM4 memory and should offer up to seven times the token throughput of Maia 200 on a bay scale, according to the Azure blog. At Ignite, Satya Nadella lent it 40% more tokens per watt than the leading GPU systems, without naming these systems or giving a date. In October, Microsoft announced deployments of several gigawatts starting in 2027. The chip must serve the group’s products, including Copilot, GitHub Copilot and Foundry, and the announcements do not mention direct rental to customers.
Cobalt 200 should represent a significant part of Azure’s new computing capabilities from 2027. Among its first customers, Microsoft cites OpenAI.
Quotes translated from English by ActuIA.
Our articles will be prioritized in Google’s “Featured” section.

