Industry observers note that at today’s Microsoft Ignite 2026 keynote, the company introduced two upcoming frontier AI models built for what the company describes as “everyday” coding tasks. But there’s a bigger story: what these models might mean for Microsoft’s AI costs and its dependence on OpenAI.

Microsoft claims its MAI Code 2 model shows 96% of Claude Opus 5.5 performance at 5% of Opus 5.5’s price.

Also: Microsoft wants to put a dollar value on AI’s work – but can you trust the math?

According to recent reports, microsoft indicates MAI Code 2 offers comparable performance to GPT-6.1 Sol or Opus 5.5 at a fraction of the cost. Microsoft’s Code 2 Flash model is comparable to the lower-end GPT-6 Luna and Claude Haiku 5.5, also at considerably lower cost.

Sources indicate that this doesn’t mean that those of us who pay for $100/month subscriptions will suddenly get the equivalent for five bucks. Instead, these are usage-based metrics based on API usage.

It’s also crucial to understand that the benchmarks are all Microsoft-provided. Performance numbers may well differ once the models undergo independent performance tests.

Throughout Satya Nadella’s keynote, Microsoft showed slides referencing a wide range of foundation models that are intended to work with Copilot and other Microsoft offerings. Microsoft famously 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.

But now, Redmond is benchmarking its internally built models against Anthropic and OpenAI offerings, and also touting that it’s integrating these latest models as optional choices for GitHub Copilot, Microsoft 365, and Fabric.

Also: 75% of developers I surveyed prefer Claude Code over Codex.

One thing that Nadella made clear in today’s keynote is how central AI is becoming in everything Microsoft does. Nearly all of those AI operations previously involved the tech giant paying some token fees back to the frontier model providers or customers doing the same thing.

But with high-performing models of its own, Microsoft may become substantially less dependent on outside model providers, especially for the incredible volume of fairly basic and routine AI requests that occur across the tech giant’s vast product line.

Recent findings suggest that if you think back over the history of agentic work models like ChatGPT Work and Claude Cowork, you may recall that they originated as agentic coding tools. As customers adopted the coding harnesses in Claude Code and OpenAI Codex to off-label productivity applications, the AI companies noticed and packaged work-centric agentic tools that were more non-programmer-friendly.

Also: I let ChatGPT Work and Claude Cowork loose on my files – only one made me nervous.

In a recent development, there’s a lot of similarity between a coding agent and a work agent, including understanding a complex environment, using tools to change things, and producing reliable results. Coding turns out to be a very good environment for developing and testing these capabilities, particularly because code output is highly measurable. Coding agents are effective training grounds for non-coding work agents.

This becomes really interesting in today’s announcement. Along with Code 2 and Code 2 Flash, Microsoft confirmed MAI Work Flash, which it describes as having been derived from Code 2 Flash. Work Flash can help generate formulas in Excel, work with documents and instructions in Word, track decisions and extract useful information from Teams, and more.

Recent findings suggest that think about the level of intelligence these tasks are demanding from the AI models. Not that much. As long as the models are good enough, they’ll be good enough.

If Microsoft can reduce its expenses and use reliable models at a much lower price, the firm itself will be able to save quite a bit of money across its entire software economy, while at the same time maintaining a deeper level of vertical integration and, therefore, control.

According to recent reports, let’s run some hypothetical numbers. Let’s say a company is spending $0.01 per AI request, which, with these new models, can drop down to $0.001 per request.

In a recent development, also: Want to try ChatGPT Wallet? Read OpenAI’s liability disclaimer very carefully first.

A company with a daily inference cost of a million dollars (a company like Microsoft, for example) could drop that daily cost to about $100,000. Add that up across a year, and the change in models could save $325 million annually.

Recent findings suggest that A million here, a million there. Pretty soon, you’re talking real money.

There’s no direct evidence that Microsoft is trying to produce replacement models for the frontier providers’ offerings. But it is an interesting and compelling story if we look at Microsoft developing its own engines for the software it sells.

Microsoft notes the numbers it’s reporting at are using a 50% introductory discount. That’s why I measured against $0.001 per request instead of $0.0005 per request above.

Sources indicate that benchmark selection could be cherry-picking. The numbers are impressive. But as we all know, there are lies, damned lies, and benchmark results. I’m fairly sure Microsoft isn’t padding these numbers too much, because it benefits from the newly introduced models more than even its own customers, but still. It’s an open glitch.

The organization also described clearly that these latest models are not designed to handle the most challenging agentic and research workloads. You’re probably not going to use them to solve a decades-old mathematics challenge. But you’re very likely to use them to reformat your PowerPoint.

Speaking of reformatting your PowerPoint, Microsoft also emphasized that these latest models did unusually well on visualization assignments. That’s a very good fit for not only UI design challenges but also graphics and slide production.

→ Read original article