Models

Microsoft's MAI Code 2 hits 96% of Claude Opus 5.5 at 5% of the cost

Microsoft says MAI Code 2 hits 96% of Claude Opus 5.5's performance at 5% of the price — a claim that could save it hundreds of millions and loosen its OpenAI dependence.

Microsoft says its new AI coding model rivals Claude Opus 5.5 at 5% of the cost
Microsoft says its new AI coding model rivals Claude Opus 5.5 at 5% of the costAMONWAT DUMKRUT / unsplash
By Sophie Lindqvist5 min read

Why it matters

  • Microsoft claims MAI Code 2 delivers 96% of Claude Opus 5.5 performance at 5% of the price.
  • Three new models debuted at Ignite 2026: MAI Code 2, Code 2 Flash and MAI Work Flash.
  • A hypothetical $1M daily inference cost could drop to ~$100,000, saving ~$325M annually.
  • Reported pricing includes a 50% introductory discount.
  • The models are integrated as options in GitHub Copilot, Microsoft 365 and Fabric.

Microsoft claims its new MAI Code 2 model delivers 96% of Claude Opus 5.5's performance at 5% of Opus 5.5's price. The company unveiled the model, along with Code 2 Flash and MAI Work Flash, at its Ignite 2026 keynote, and the pricing claim is based on API usage metrics rather than subscription tiers.

That distinction matters. If you pay $100 a month for a coding subscription, you won't suddenly get the same thing for five bucks. What Microsoft is describing is the per-request economics that apply when AI workloads run at scale — the kind of scale Microsoft itself operates at across Copilot, Microsoft 365 and Fabric.

What did Microsoft actually announce?

Three models, two of them frontier-class and aimed at what the company calls "everyday" coding tasks. Here's the lineup:

  • MAI Code 2 — Microsoft says it offers comparable performance to GPT-6.1 Sol or Claude Opus 5.5 at a fraction of the cost.
  • Code 2 Flash — positioned against the lower-end GPT-6 Luna and Claude Haiku 5.5, again at considerably lower cost.
  • MAI Work Flash — a work-focused agent derived from Code 2 Flash.

Satya Nadella's keynote slides referenced a wide range of foundation models intended to work with Copilot and other Microsoft offerings. The new models are being integrated as optional choices inside GitHub Copilot, Microsoft 365 and Fabric.

The stakes here go beyond raw benchmarks. Microsoft has a multi-billion-dollar investment in OpenAI dating back to 2023 and has built many of its products and services around OpenAI intellectual property. Now Redmond is benchmarking its internally built models directly against Anthropic and OpenAI offerings — a clear signal that it's building its own engine room.

Why does this matter for Microsoft's OpenAI dependence?

Nadella made clear in the keynote how central AI has become to everything Microsoft does. Nearly all of those AI operations previously involved Microsoft paying token fees back to frontier model providers, or customers paying the same way.

With high-performing models of its own, Microsoft could become substantially less dependent on outside providers. That's especially true for the enormous volume of fairly basic, routine AI requests that flow across the company's vast product line every day.

Consider what those requests actually demand. Generating an Excel formula. Extracting decisions from a Teams transcript. Reformatting a slide. These don't need frontier-level intelligence. As the source puts it: as long as the models are good enough, they'll be good enough.

The result is deeper vertical integration — and more control — over the software stack Microsoft sells.

How much money are we talking about?

Run the hypothetical math and the numbers get large fast. Suppose a company pays $0.01 per AI request, and the new models bring that down to $0.001 per request.

A company spending $1 million a day on inference — a company like Microsoft, for instance — could see that drop to roughly $100,000 a day. Over a full year, that's about $325 million in savings. As the source dryly notes: a million here, a million there, and pretty soon you're talking real money.

There's no direct evidence Microsoft is building drop-in replacements for everything OpenAI and Anthropic sell. But the pattern is compelling: Microsoft developing its own engines for the software economy it dominates.

Where did MAI Work Flash come from?

The Office-focused model has an interesting lineage, and it traces back through the history of agentic tools. Platforms like ChatGPT Work and Claude Cowork originated as agentic coding tools. When users began pushing coding harnesses like Claude Code and OpenAI Codex into off-label productivity work, the AI companies noticed and packaged work-centric agents that non-programmers could actually use.

The two jobs share a lot of machinery. Both require understanding a complex environment, using tools to change things, and producing reliable results. Coding turns out to be an excellent training ground for those capabilities, largely because code output is highly measurable.

Microsoft carried that logic into today's announcement. MAI Work Flash, derived from Code 2 Flash, can generate formulas in Excel, work with documents and instructions in Word, and track decisions and extract useful information from Teams.

The takeaway: coding agents double as a proving ground for the office agents Microsoft is now shipping.

Can you trust the benchmark numbers?

Microsoft's claims look impressive, but several caveats are worth holding onto.

First, the pricing. Microsoft said the reported numbers include a 50% introductory discount — which is why the source's math used $0.001 per request rather than $0.0005.

Second, benchmark selection. The numbers are strong, but benchmark results can flatter whoever picks them. The source's assessment is measured: "I'm fairly sure Microsoft isn't padding these numbers too much, because it benefits from the new models more than even its own customers, but still. It's an open issue."

Third, scope. Microsoft stated clearly that these models are not designed for the most challenging agentic and research workloads. You probably won't use them to solve a decades-old mathematics problem. You will very likely use them to reformat your PowerPoint.

One genuinely notable result: Microsoft emphasized that the new models performed unusually well on visualization assignments. That's a strong fit for UI design work, graphics and slide production — tasks that sit squarely in the middle of Microsoft's productivity empire.

The open question is whether the 96%-at-5% claim holds up under independent testing once the introductory pricing fades. If it does, Microsoft's quiet march toward model self-sufficiency just became a lot harder for OpenAI and Anthropic to ignore.

Original: microsoft.ai

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.

749 articles

Related articles

  1. Microsoft AI Launches MAI Code 2 and Flash on Cost-per-Task Math
  2. Microsoft’s Ignite 2026 adds Copilot Workspaces and agent controls
  3. Meta and Microsoft sharply cut Claude usage as Anthropic becomes a rival
  4. Microsoft's AI Numbers Tell Two Stories: Growth and Strain
  5. Anthropic's Claude Opus 5.5 costs 40% less, hits Fable 5.1 performance

« Previous articleNext article »