Skip to content
AI EmployeeSuper-AgentsAgent-to-AgentBenchmarksPricingBlog

Microsoft's New MAI Models: Code 2 vs Claude Opus 5.5

At a glanceQuick answers
Which models did Microsoft launch?
Seven, at Ignite on November 17, 2026: MAI Code 2, Code 2 Flash, Cyber 2 Flash, Image 2.7, Realtime 1, MAI-Transcribe-2 and Decision 1.1.
How does MAI Code 2 compare with Claude Opus 5.5?
Microsoft says 96% of Opus 5.5’s performance at 5% of the cost, on its own basket of coding benchmarks.
Are the figures independent?
No. They are Microsoft’s own claims, made on launch day.
Editorial data illustration on a near-white ground under the words MAI Code 2 vs Claude Opus 5.5: a tall charcoal coin stack labelled Claude Opus 5.5 beside a tiny orange stack labelled MAI Code 2, with 96% of the performance and 5% of the cost in large type.
Fig 0Microsoft's headline claim for MAI Code 2, drawn from its own figures. Made by CellCog's image agent, running GPT Image 2.6.

Microsoft AI launched seven new models at Ignite on November 17, 2026, and priced its pitch against Anthropic and OpenAI. The headline is MAI Code 2, which Microsoft says gives 96% of Claude Opus 5.5’s coding performance at 5% of the cost. The launch post on X read: “Frontier performance. A fraction of today’s prices. Built to run 24/7.”

Every figure on this page is Microsoft’s own, from its posts on launch day. Microsoft also says plainly what the models are not built for, and that line matters as much as the prices. This page records what Microsoft published on November 17, 2026.

On this page · 7 sectionsOpen
  1. What Microsoft launched
  2. The cost claims
  3. What the benchmarks are, and what the models are not for
  4. Licensed weights and Decision 1.1
  5. What to check before switching
  6. Where CellCog sits
  7. Sources
Key points5 · 7 min full read
  1. Seven rounded blocks in a row, the last one orange.
    Microsoft AI launched seven models at Ignite on November 17, 2026: MAI Code 2, Code 2 Flash, Cyber 2 Flash, Image 2.7, Realtime 1, MAI-Transcribe-2 and Decision 1.1.
  2. A code symbol with an orange price tag.
    Microsoft says MAI Code 2 gives 96% of Claude Opus 5.5’s coding performance at 5% of the cost, and 98% of GPT-6.1 Sol’s at 40%.
  3. A shield with an orange target on it.
    MAI Cyber 2 Flash, Microsoft says, beats GPT-6 Astra on the CyberGymE2E benchmark at less than 2% of the cost per task.
  4. A mountain with a path that stops at an orange flag.
    Microsoft says these models are not yet tuned for fundamental science research or very long-horizon agentic work.
  5. An orange key in front of a cloud.
    MAI Licensed Weights lets a customer tune a frontier MAI model on its own data and run it in its own Azure environment.

§ 01What Microsoft launched

Microsoft AI posted the launch at 18:41 UTC, and listed the models an hour later: “We urgently need secure, aligned, and contained frontier intelligence at a fraction of today’s prices.” It had about 18,000 views when we read it at 20:12 UTC.

Model What it is for Microsoft’s claim
MAI Code 2 Repo-wide reasoning 98% of GPT-6.1 Sol’s performance at 40% of the cost; 96% of Claude Opus 5.5’s at 5%
MAI Code 2 Flash Fast, high-volume coding Matches GPT-6 Luna and Claude Haiku 5.5 on everyday coding at 65% and 20% of the price
MAI Cyber 2 Flash Security work Beats GPT-6 Astra on CyberGymE2E at less than 2% of the cost per task
MAI Image 2.7 Images #1 on Artificial Analysis image editing, ahead of GPT Image 2.6 Sunburst (max) at 20% of the price
MAI Realtime 1 Real-time audio Beats GPT-Realtime 2.1, from $0.01 per audio minute
MAI-Transcribe-2 Transcription, behind Dragon Copilot Transcribe Powers 10 million patient encounters a month and halves multilingual error rates
Decision 1.1 Decisions inside applications Beats Jev, Google and OpenAI on the Artificial Analysis Decision Index
Table 1The seven MAI models, as Microsoft described them on November 17, 2026

§ 02The cost claims

The coding claim is the one aimed at the model most developers compare against. In Microsoft’s words: “MAI Code 2 gives you 98% of GPT-6.1 Sol’s performance at 40% of the cost, and 96% of Claude Opus 5.5’s at just 5% of the cost.” And for the smaller model: “MAI Code 2 Flash matches GPT-6 Luna and Claude Haiku 5.5 on everyday coding tasks, at 65% and 20% of the price.”

Comparison Cost or price, % of the rival’s
MAI Code 2 vs GPT-6.1 Sol 40
MAI Code 2 vs Claude Opus 5.5 5
MAI Code 2 Flash vs GPT-6 Luna 65
MAI Code 2 Flash vs Claude Haiku 5.5 20
MAI Image 2.7 vs GPT Image 2.6 Sunburst (max) 20
Table 2What Microsoft says each MAI model costs, as a share of the rival’s
Cost of an MAI model as a share of the rival's, in percent (Microsoft's figures)Bar chart of the cost or price of an MAI model as a share of its rival's, in percent, from Microsoft's posts: MAI Code 2 vs GPT-6.1 Sol 40; MAI Code 2 vs Claude Opus 5.5 5; MAI Code 2 Flash vs GPT-6 Luna 65; MAI Code 2 Flash vs Claude Haiku 5.5 20; MAI Image 2.7 vs GPT Image 2.6 Sunburst (max) 20.Code 2 vs GPT-6.1 Sol40Code 2 vs Claude Opus 5.55Code 2 Flash vs GPT-6 Luna65Code 2 Flash vs Claude Haiku 5.520Image 2.7 vs GPT Image 2.6 Sunburst (max)20Cost of an MAI model as a share of the rival's, in percent (Microsoft's figures)Bar chart of the cost or price of an MAI model as a share of its rival's, in percent, from Microsoft's posts: MAI Code 2 vs GPT-6.1 Sol 40; MAI Code 2 vs Claude Opus 5.5 5; MAI Code 2 Flash vs GPT-6 Luna 65; MAI Code 2 Flash vs Claude Haiku 5.5 20; MAI Image 2.7 vs GPT Image 2.6 Sunburst (max) 20.Code 2 vs GPT-6.1 Sol40Code 2 vs Claude Opus 5.55Code 2 Flash vs GPT-6 Luna65Code 2 Flash vs Claude Haiku 5.520Image 2.7 vs GPT Image 2.6 Sunburst (max)20
Fig 1Cost of an MAI model as a share of the rival's, in percent (Microsoft's figures)

The security model gets its own claim: less than 2% of GPT-6 Astra’s cost per task on CyberGymE2E, while beating it. Microsoft AI framed that one as a matter of running agents all day: “Our national cyber defense will fall behind if we can’t run security agents 24/7.”

Animation: a grey block labelled Claude Opus 5.5 and a smaller orange block labelled MAI Code 2 take the same coding task; their score bars rise almost level, then twenty coins stack under Claude Opus 5.5 and one coin under MAI Code 2
Fig 2Microsoft says MAI Code 2 scores 96% of Claude Opus 5.5 at 5% of the cost

§ 03What the benchmarks are, and what the models are not for

  • The coding basket. For MAI Code 2, Microsoft averages DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, which it says best reflect the usage it sees in GitHub Copilot. For MAI Code 2 Flash, it averages the first three.
  • The limit, in Microsoft’s words. The models are built for everyday coding work, and “they’re explicitly not yet tuned for harder problems, like fundamental science research and very long-horizon agentic work.”
  • The argument. Microsoft’s case for a cheaper model is that frontier models carry a cost on every call: “When the same model is built to also push the frontier of science, you pay for that capability on every task, even for a routine bug fix.”

§ 04Licensed weights and Decision 1.1

  • MAI Licensed Weights. A customer can take a frontier MAI model and tune it on its own data. Microsoft’s description: “Take a frontier MAI model, customize it using your proprietary data if you want to. Then run it securely in your own Azure environment.”
  • Decision 1.1. The successor to the decision model Microsoft released in October, which we covered in Microsoft-Decision-1: Price, Benchmarks, vs Jev. Microsoft says 1.1 beats Jev, Google and OpenAI on the Artificial Analysis Decision Index.
  • Earlier MAI work. Microsoft’s base model is covered in Microsoft MAI-Base-2.

§ 05What to check before switching

Our reading, not Microsoft’s: the percentages compare cost at a given level of performance on benchmarks Microsoft picked, on the day it launched. Before moving coding work to MAI Code 2, three things are worth checking on your own tasks.

  1. The 4% gap. 96% of Claude Opus 5.5’s score on an average can hide larger gaps on single tasks. Run your hardest recent tickets through both.
  2. Long tasks. Microsoft says these models are not yet tuned for very long-horizon agentic work. Agents that run for hours on one goal are exactly that.
  3. Price per finished task. A cheaper model that needs more retries or more review can cost more per merged change. Count retries and review time, not only the token bill.

§ 06Where CellCog sits

Our conflict, declared: we build CellCog, a platform where a business hires AI employees, and those employees run on Anthropic’s Claude models, Claude Opus 5.5 at most tiers and Claude Fable 5.2 at Agent Max and Creative Max. That is the model Microsoft’s headline compares against. CellCog does not run on any MAI model, and an AI employee’s work is mostly the long, many-step kind that Microsoft says these models are not yet tuned for.

§ 07Sources

Frequently asked5 questions

Q1What did Microsoft launch at Ignite 2026?

Seven Microsoft AI models on November 17, 2026: MAI Code 2, MAI Code 2 Flash, MAI Cyber 2 Flash, MAI Image 2.7, MAI Realtime 1, MAI-Transcribe-2 and Decision 1.1, plus MAI Licensed Weights for running a tuned model in your own Azure environment.

Q2How does MAI Code 2 compare with Claude Opus 5.5?

Microsoft says MAI Code 2 reaches 96% of Claude Opus 5.5’s performance at 5% of the cost, and 98% of GPT-6.1 Sol’s at 40%. These are Microsoft’s figures on a basket of coding benchmarks it chose.

Q3Which benchmarks are behind the coding claims?

For MAI Code 2, an average of DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. For MAI Code 2 Flash, an average of the first three.

Q4What does MAI Realtime 1 cost?

Microsoft says it starts at $0.01 per audio minute and outperforms GPT-Realtime 2.1.

Q5Does CellCog use MAI models?

No. CellCog’s AI employees run on Anthropic’s Claude models: Claude Opus 5.5 at most tiers and Claude Fable 5.2 at Agent Max and Creative Max.

Published 17 November 2026 All Choosing a platform →