MAI Code 2 claims 96% of Opus 5.5’s coding score at 5% of the cost
Microsoft's MAI Code 2 and MAI Code 2 Flash arrive in GitHub Copilot, with Microsoft claiming near-frontier coding scores at a small share of rival models' cost.

Microsoft has added two in-house coding models, MAI Code 2 and MAI Code 2 Flash, to GitHub Copilot, GitHub said in its changelog. Microsoft’s own pitch is price: 96% of Claude Opus 5.5’s benchmark performance at 5% of its cost.
Two models, split by task size
GitHub describes MAI Code 2 as a general-purpose model for deep debugging, multi-file implementation, refactoring and reasoning across a whole repository. MAI Code 2 Flash is the lighter, cheaper sibling, meant for code explanations, targeted edits, routine fixes and fast iteration. Both are built for agentic coding workflows, and developers choose between them in the model picker.

MAI Code 2 is offered on Copilot Pro, Pro+, Max, Business and Enterprise. MAI Code 2 Flash also reaches Copilot Free and Student. GitHub says both are open to eligible subscribers on annual plans. The picker is available in Visual Studio Code, Visual Studio, Copilot CLI, the Copilot cloud agent, the Copilot app, github.com, GitHub Mobile on iOS and Android, JetBrains, Xcode and Eclipse.
Both models are billed under usage-based billing. The changelog gives no per-token price and points to GitHub’s models and pricing page. On Business and Enterprise plans, administrators control access through the model policy in Copilot settings. Under default enablement, new models switch on automatically unless the global default is off or the model is explicitly disabled.

Microsoft’s cost-per-score claims
The performance figures come from Microsoft AI’s own announcement, not from GitHub. Microsoft averaged results over four benchmarks: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. On that average it says MAI Code 2 delivers 98% of GPT 6.1 Sol’s performance at 40% of its cost, and 96% of Claude Opus 5.5’s performance at 5% of its cost. In blind visual frontend tests with 258 annotators, reviewers preferred MAI Code 2 over GPT-6 Sol 64% of the time, according to Microsoft.
For Flash, Microsoft says it matches GPT-6 Luna and Claude Haiku 5.5 on three everyday coding benchmarks, at 65% and 20% of their cost respectively. The company says its benchmarks ran at maximum reasoning effort. It adds that results depend on workload, configuration and pricing assumptions, and that the costs shown reflect a 50% promotional price.

What the cost numbers do and do not show
The headline ratios compare cost for a similar score, so they say little about the absolute bill. Since the 50% promotional price is built into Microsoft’s cost figures, the gap could narrow when standard pricing applies. Microsoft publishes no per-token rates in the material we read, so the claim cannot be checked against a price list, and the benchmark results are the company’s own and not independently verified.
The pricing question matters because Copilot is moving to usage-based billing, where a cheaper model for the same task lowers the user’s bill directly. A model that sits near Opus-class scores at a small share of the cost gives teams a reason to route routine agent runs away from the priciest options. Microsoft itself says MAI Code 2 is also becoming the workhorse behind its own products, so volume economics clearly matter to the company.
The same week, GitHub reported that Copilot Fusion Max reached 71.2% on Terminal-Bench 4.1, which it said cost 88% less than Fable 5.2 and 57% less than Opus 5.5, as we covered in our report on the Fusion leaderboard result. GitHub’s Fusion post also lists MAI-Code-2 among the models added to Fusion Standard. Cheap, capable models are increasingly the building blocks of those routed systems.
Limits Microsoft states itself
Microsoft says neither model is yet tuned for fundamental science research or for very long-horizon agentic work. That is a notable caveat for a release marketed around agentic coding: the target is everyday enterprise tasks such as production code, debugging and UI work, not multi-day autonomous projects.
Several details are missing. The sources we read do not state a context window, parameter counts or absolute prices. GitHub’s changelog mentions vision support but gives no specifics. Beyond Copilot, Microsoft says Flash is in limited-access preview in Microsoft Foundry, with releases planned on OpenRouter, Vercel, Hermes, Open Code and Kilo. The Copilot rollout is part of a broader Ignite push that also included changes to Copilot code review.
For developers on a paid annual plan, the practical step is simple: open the model picker and try MAI Code 2 on a routine task next to the model currently in use, comparing both result quality and usage charges.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Leo MartinsAI Tools, Prompts & Practical AI ExpertLeo Martins is the AI Tools & Practical AI Expert at Aiholics, focused on helping readers use artificial intelligence to improve productivity, creativity and everyday work.



