Microsoft Launches Six Specialized MAI Models to Challenge Rivals on AI Cost
Key Takeaways
- •Microsoft introduced or broadened access to six specialized models: Code 2, Code 2 Flash, Cyber 2 Flash, Image 2.7, Transcribe 2 and Realtime 1.
- •Microsoft reported that Code 2 achieved 96% of Claude Opus 5.5’s performance at 5% of its cost per task, and 98% of GPT-6.1 Sol’s performance at 40% of its cost.
- •Cyber 2 Flash recorded a 92.3% score on the CyberGym-E2E benchmark, compared with 91.5% for GPT-6 Astra, according to Microsoft.
- •Code 2 and Code 2 Flash promotional prices are scheduled to double on May 17, 2027.
- •Independent testing has not verified Microsoft’s comparisons, while customized MAI models are required to operate on Azure.

Microsoft showcased six MAI models at Ignite, the company's annual conference for enterprise developers and partners, on November 17. Some were newly launched, while others became more widely available. The company says its models can nearly match more expensive rivals on the tasks they are built for while costing far less, with the releases detailed in Microsoft's Ignite 2026 blog post.
Specialized models aimed at the cost of a task
Microsoft's six MAI specialized models are Code 2, Code 2 Flash, Cyber 2 Flash, Image 2.7, Transcribe 2, and Realtime 1.
Central to Microsoft's claims of lower cost at almost the same or better performance is its MAI Code 2 model, the flagship of the new coding lineup. The company tested the model using four coding benchmarks: DeepSWE 1.1, Terminal-Bench 2.1, SWE Atlas, and FrontierCode 1.1. In results published on November 17, Microsoft says Code 2 delivered 96% of Claude Opus 5.5's performance while costing just 5% as much per task. The company also compared Code 2 with GPT-6.1 Sol, reporting 98% of the performance at 40% of the cost.
As stated on GitHub's pricing page, MAI Code 2 is priced at $0.50 for every million input tokens and $2.50 for every million output tokens. The MAI Code 2 Flash version is available at a lower price of $0.05 for input tokens and $0.25 for output tokens. Claude Opus 5.5 charges $4 for input and $20 for output, while GPT-6.1 Sol charges $2 and $10, respectively. However, the prices of Microsoft's models are only promotional in nature: the costs of MAI Code 2 and MAI Code 2 Flash will double after May 17, 2027, giving enterprise buyers a fixed window to weigh the claimed cost advantages at promotional rates before full pricing applies.
Microsoft also reported its model performing better in terms of cybersecurity. Its Cyber 2 Flash achieved a 92.3% cybersecurity score on the CyberGym-E2E benchmark, compared to GPT-6 Astra, which got a score of 91.5%. The input/output rates for Cyber 2 Flash are $0.10/$0.50 per every million tokens.
The claims about the models' performance should still be taken with caution, however. Microsoft chose the benchmarks used to compare its models with rivals, and independent testing has yet to confirm the results. Until such outside checks arrive, the cost-per-task figures rest on vendor-reported measurements alone.
Suleyman frames the tradeoff
Mustafa Suleyman, Microsoft's AI chief, explained the tradeoff in an Axios interview on November 17. Models developed for certain tasks, he elaborated, have greater capability and knowledge using fewer parameters. However, there is still a limit to this specialization: a model that performs well at basic coding can run into problems on advanced science problems.
To address the concern, Suleyman said his team is trying to build “true superintelligence” — an AI system capable of performing any intellectual task better than a human being. He did not say when such a system might be available.
Copilot, Azure weights and platform control
Code 2 is already available in GitHub Copilot, Microsoft's coding assistant, which could reduce purchases of outside models. Enterprise customers can also license MAI weights to customize the models.
There is a condition attached, however. Axios reported that customized models must run on Azure, Microsoft's cloud computing platform. Lower model costs could therefore come with greater cloud dependence.
A continuation of Nadella's supplier worry
Microsoft had already introduced seven in-house models on June 2, and the latest releases also address a concern raised by CEO Satya Nadella earlier this year. According to a Cryptopolitan report on July 17, Nadella raised doubts about the idea of depending on a limited set of leading AI companies for token capacity. By creating its own models, Microsoft is able to secure ample bargaining power when negotiating with such companies.
Falling prices, uneven access
Microsoft is increasing its portfolio of AI offerings as prices across the industry continue to decrease. According to a study by Epoch AI published on September 22, it has become cheaper to achieve similar AI results since 2023 by as much as 47% quarterly. This translates to approximately 13 times lower prices every year.
Despite falling costs, established AI providers still dominate business spending. Ramp's October 2026 index shows Anthropic accounted for 49.3% of token spending among US businesses using its platform, with OpenAI following at 46.1% for the week ending November 1. The figures do not represent the global market. Future updates to such spending trackers will offer one reading on whether lower-priced in-house models begin to shift enterprise token budgets.
AI infrastructure is also heavily concentrated in a few countries and companies. Stanford's 2026 AI Index counted 5,427 data centers in the United States in 2025 — more than ten times the number in any other country. Stanford also found that Taiwan's TSMC manufactures almost every leading AI chip.
The concentration has drawn regulatory attention. A July 10 OECD report warned that control over chips and cloud infrastructure could make it harder for other companies to compete in the AI market.
Microsoft's new MAI models raise a similar concern. Lower prices could make its AI models more affordable for businesses worldwide to adopt, but the Azure requirement could also make it harder for customers to switch cloud providers, even if they save money on the models themselves.