Microsoft Unveils 7 MAI Models, Targets OpenAI Pricing
At Ignite, Microsoft introduced seven proprietary models, claiming MAI Code 2 achieves 96% of Claude Opus 5.5's performance at 5% of the cost. Comparison with GPT-6.1 Sol relies on promotional pricing until May 2027.

Microsoft unveiled seven proprietary MAI models developed by the Microsoft AI division led by Mustafa Suleyman on Tuesday (17) at the kickoff of Ignite in San Francisco. The package covers coding, security, image generation, real-time voice, transcription, and model routing, with a central message focused on pricing. According to Microsoft, MAI Code 2 delivers 96% of the performance of Claude Opus 5.5 from Anthropic for about 5% of the cost.
The figures are drawn from the company's own charts. In its benchmark test, MAI Code 2 scored 63.2%, compared to 64.1% for OpenAI's GPT-6.1 Sol and 65.3% for Opus 5.5. The average cost per task stood at $0.309, against $0.761 for Sol and $6.306 for Opus. Media outlets covering the launch have yet to find independent evaluations that confirm these comparisons.
The Price Asterisk
The GitHub pricing page lists MAI Code 2 at $0.50 per million input tokens and $2.50 per million output tokens. This is a promotional price, with a 50% discount valid until May 17, 2027. After that date, the price doubles to $1.00 and $5.00. The Flash version is also on promotion at $0.05 and $0.25.
This detail alters the interpretation of one of the most frequently repeated phrases at the event. Suleyman presented the model as capable of performing the tasks of GPT-6.1 Sol for 40% of the price, but Microsoft's footnote clarifies that this calculation uses the promotional price. According to FourWeekMBA's calculations, at full price, the relationship rises to nearly 80% of Sol's cost.
The comparison with Opus withstands the asterisk better, as the starting differential is of an order of magnitude. Even with the price doubled, the cost per task for MAI Code 2 would remain below 10% of what Anthropic's model recorded in Microsoft's own test.
From Copilot to SOC
MAI Code 2 is already available in paid plans of GitHub Copilot. The Flash version is also available in free and student plans, along with a limited preview in Microsoft Foundry. For Microsoft, the distribution channel is as important as the model: Copilot is where millions of developers already choose which model responds to their next request, and now the menu's cheapest option comes from the house.
In security, MAI Cyber 2 Flash has been reserved for verified defenders, an access restriction that acknowledges the dual-use of a model trained to find and exploit vulnerabilities. MAI Realtime 1, the company's first voice-to-voice real-time model, starts at $0.01 per minute of audio. The Decision-1.1 routing model is now generally available.
The Relationship with OpenAI
The launch coincides with OpenAI unveiling GPT-6.2 Astra and GPT-6.2 Sol, models that Microsoft also distributes on Azure. The Redmond company now sells both its partner’s models, in which it has invested billions, and its own alternatives explicitly positioned as cheaper. For corporate buyers, this presents bargaining power: Microsoft has an incentive to push its own model, which does not share revenue, and customers can leverage this difference during contract renewal.
A skeptical reading also deserves consideration. Benchmarks chosen by the vendor rarely replicate the legacy code of a bank or an insurance company, and the cost per task measured in a lab ignores rework, human review, and failures in agent chains. A model that hits 63% and costs one-twentieth may end up being more expensive than one that hits 65% if the error occurs in a critical part of the system.
Where the Numbers Matter
In the United States, the immediate effect appears in the budget for Copilot seats and the token consumption of engineering teams. In India, the channel aligns with the pricing model of integrators: fixed-price development contracts executed in delivery centers in Bangalore, Pune, and Hyderabad now include the cost of the code agent, and a 20-fold cheaper input pressures the fees charged per developer. In Brazil, where banks and retailers outsource much of their legacy system maintenance to local and Indian consultancies, the same calculation arises at the purchasing table when the supplier has to explain why maintenance hours haven’t dropped.
The real test occurs in May 2027. If adoption of MAI Code 2 persists beyond the promotional period, Microsoft will have demonstrated that it can compete on model, not just distribution. If usage reverts when the price doubles, Ignite will have been a subsidized customer acquisition campaign.
Sources
- blogs.microsoft.comhttps://blogs.microsoft.com/blog/2026/11/17/ignite-2026-building-the-frontier-ecosystem/
- testingcatalog.comhttps://www.testingcatalog.com/microsoft-launches-seven-mai-models-at-ignite-2026/
- mlq.aihttps://mlq.ai/news/microsoft-launches-seven-mai-models-at-ignite-pitching-its-coding-model-at-5-of-claude-opuss-price/
- fourweekmba.comhttps://fourweekmba.com/ai-microsofts-mai-code-2-costs-40-of-sol-at-its-promo-price/
- siliconangle.comhttps://siliconangle.com/2026/11/17/microsoft-rolls-out-new-first-party-models-and-updates-to-agentic-ai-continuous-improvement/