Microsoft Unveils MAI Code 2 to Rival Claude Opus
At Microsoft Ignite on , Microsoft unveiled two in-house frontier models, MAI Code 2 and MAI Code 2 Flash, along with a workplace agent called MAI Work Flash. CEO Satya Nadella said the flagship MAI Code 2 delivers 96% of Claude Opus 5.5’s performance at roughly 5% of the token cost.

The release is aimed at enterprise API infrastructure and code execution costs across Microsoft’s commercial products. It does not touch consumer Copilot subscription tiers. Instead, it gives Microsoft a cheaper internal alternative to frontier models from outside partners.

The 5% Figure Comes With a Discount Attached

By Microsoft’s own benchmarks, MAI Code 2 matches capabilities previously led by GPT-6.1 Sol and Claude Opus 5.5. The smaller MAI Code 2 Flash targets parity with latency-focused systems such as GPT-6 Luna and Claude Haiku 5.5.

Microsoft ran its cost comparisons using a 50% introductory API discount, so the savings will shift once base pricing takes effect. Independent evaluations have not yet confirmed any of the company’s numbers, and Microsoft acknowledged the models are strongest at structural visualization and everyday programming rather than frontier mathematics or speculative reasoning.

The models are being built into GitHub Copilot and Fabric, which lets developers skip third-party token meters on standard software pipelines.

MAI Work Flash Takes the Same Base Into Microsoft 365

MAI Work Flash is an enterprise agent model derived directly from the Code 2 Flash architecture. Microsoft’s reasoning is that code harnesses test precise, verifiable actions, so structured software logic made a sound training foundation for automated office tasks.

The model runs natively across Microsoft 365 apps and targets work that needs low latency without heavy compute:

  • Summarizing meeting decisions and action items in Teams
  • Parsing structured syntax and writing calculation formulas in Excel
  • Drafting standard executive memos and technical documentation in Word
  • Running routine data pipelines through Microsoft Fabric

The motive is easy to read. Microsoft has held a multi-billion-dollar investment in OpenAI since 2023, yet every routine Copilot request routed to an outside provider carries a third-party token markup. Handling those requests on its own models removes that cost.

Hashlytics Take

Strip away the headline ratio and this is a pricing lever. The 96% figure is Microsoft’s own, and the 5% cost figure was calculated with an introductory 50% discount, so the real gap narrows once list prices apply. Even so, a model that handles everyday coding and Office tasks competently at a fraction of frontier pricing changes the negotiation with every outside supplier, OpenAI included. The question to watch is whether routine traffic in GitHub Copilot and Fabric defaults to MAI models, because that routing decision, not a benchmark chart, determines who gets paid when a Copilot request runs. Until independent evaluations land, treat the parity claim as a sales number.

Follow Hashlytics on Bluesky, Facebook, LinkedIn , Telegram and X to Get Instant Updates