Microsoft claims MAI Code 2 delivers 96% of Claude Opus 5.5 performance at 5% of the price

Microsoft launched a new offering on Tuesday that lets enterprise customers customize the company's own MAI models with their own data and run them in their own Azure environment.

File photo: Building 92 at Microsoft Corporation headquarters in Redmond, Washington. Photographed by user Coolcaesar on 30 May 2016.
File photo: Coolcaesar, CC BY-SA 4.0 · source This adaptation is licensed under CC BY-SA 4.0.

Microsoft claims MAI Code 2 delivers 96% of Claude Opus 5.5 performance at 5% of the price

Microsoft launched a new offering on Tuesday that lets enterprise customers customize the company's own MAI models with their own data and run them in their own Azure environment.

Microsoft announced a new offering on Tuesday that lets enterprise customers customize the company's own MAI models with their own data and run them in their own Azure environment. The announcement came at the Ignite keynote, alongside new models and a Microsoft-funded cost comparison aimed directly at ChatGPT and Claude — a clear signal that Microsoft AI is now competing with the very partners it depends on.

What was announced

The new business offering is called MAI Licensed Weights. According to GeekWire's coverage of the Ignite keynote on November 17, 2026, companies can take a Microsoft AI model, customize it with their own data, and run it in their own Azure environment, with assistance from Microsoft's consulting arm, Frontier Company (GeekWire).

The program launches alongside new models from Microsoft's own MAI family, including MAI Code 2 and MAI Cyber 2 Flash. This is not about access to third-party models, but to Microsoft's own, homegrown models — which is the whole point of the strategic line.

There is already a precedent for what this kind of customization looks like in practice. AT&T used an earlier version of the program, Custom Models, to build a MAI model for analyzing its 5G network. That illustrates the type of use case Microsoft is betting on: domain-specific analytical tasks where the customer's own data gives the model value that a general-purpose model lacks.

Self-sufficiency as strategy

According to GeekWire, MAI Licensed Weights is the latest step in Microsoft's push toward what Microsoft AI chief Mustafa Suleyman has called "long term self-sufficiency." That line began with seven homegrown models presented at the Build conference in June.

By building its own models and now letting customers customize and operate them themselves, Microsoft moves the competition into territory where it is no longer dependent on external model providers.

The Information, with Aaron Holmes as byline, was first to report on the plan, under the headline "Microsoft to Let Customers Customize Its AI Models' 'Weights'" (The Information). The article text is behind a paywall, so the details of what the program actually involves rest on GeekWire's account of the keynote.

The competitive frame: the numbers Microsoft itself supplied

The keynote was as much an attack on OpenAI and Anthropic as it was a product launch. GeekWire described the competitor framing as clearly visible throughout the presentation.

The most striking claim concerned MAI Code 2. Microsoft says the model delivers 96 percent of Claude Opus 5.5's coding performance at 5 percent of the price — measured on benchmarks Microsoft itself chose (GeekWire). That is a vendor claim, not an independent verification: the reference model, the benchmarks, and the pricing were all chosen by the party with the most to gain from the comparison.

The Copilot comparison follows the same pattern. Research from GigaOm, commissioned and paid for by Microsoft, laid out monthly costs per enterprise user based on modeled work tasks that Microsoft ran tests on in October:

  • For an "everyday user": Copilot at $30, versus $81 for ChatGPT and $205 for Claude.
  • For advanced agent users: $96 for Copilot, versus $149 for ChatGPT and $299 for Claude.

GeekWire itself notes the caveats: the figures are Microsoft-funded estimates built on modeled workloads, not real-world billing from enterprise customers. They say something about how Microsoft itself wants to position its pricing, more than they prove anything about what customers actually pay or get. The context is still meaningful: Anthropic has recently expanded into productivity tools in Office territory, and Microsoft is responding by framing its pricing squarely against those two competitors.

What we don't know

The program's name elevates the most technical term in AI — model weights, the learned parameters themselves that define what a model can do. But there is no primary Microsoft documentation of what "customize the weights" actually means for customers:

  • Depth of customization. Is this full fine-tuning of all parameters, adapters, or limited customization? The word choice suggests more than typical fine-tuning-as-a-service, but the scope is not specified.
  • Licensing and IP terms. What happens to a customized model if the customer ends the relationship? Who owns the customized weights? Not documented.
  • Which models qualify. It is unclear whether all MAI models are included, or only a selection.
  • Pricing and availability. Neither pricing for the program nor a launch date beyond the announcement is known.

These gaps are worth taking seriously because the terms will determine whether the program genuinely gives enterprises more control, or is in practice an advanced hosting service with Microsoft still managing the models.

Why it matters

Even with the caveats, three things stand. First, MAI Licensed Weights signals that Microsoft is now selling deep model control — down to the weights themselves — as an enterprise product, shifting the competition from "which model is best" to "who controls your model." Second, the cost announcements show Microsoft is willing to compete with OpenAI and Anthropic on price, not just integration, and is willing to commission research to support that message. Third, this is the commercial follow-through on the self-sufficiency strategy Suleyman began with seven of its own models in June: the strategy now has a billable form.

The open question is whether customers will interpret "their own weights in their own environment" as genuine independence — and whether Microsoft's own benchmark numbers hold up when third parties test them. Both will determine whether this becomes a marginal enterprise add-on service or a shift in who supplies the models behind the biggest enterprises.

AIMag.no
AIMag.no
The AIMag.no editorial team covers artificial intelligence, tools, research, and regulation.

Get the best of AI MAG in your inbox

News, analysis, and ideas at the intersection of AI and society.