ZDNET’s cardinal takeaways
- Microsoft unveiled 3 caller AI models astatine Ignite 2026.
- It claims 96% of Claude Opus show astatine 5% of the cost.
- Cheaper models could prevention Microsoft hundreds of millions.
At today’s Microsoft Ignite 2026 keynote, the institution introduced two caller frontier AI models built for what the institution describes arsenic “everyday” coding tasks. But there’s a bigger story: what these models mightiness mean for Microsoft’s AI costs and its dependence connected OpenAI.
Microsoft claims its MAI Code 2 exemplary shows 96% of Claude Opus 5.5 show astatine 5% of Opus 5.5’s price.
Microsoft says MAI Code 2 offers comparable show to GPT-6.1 Sol oregon Opus 5.5 astatine a fraction of the cost. Microsoft’s Code 2 Flash exemplary is comparable to the lower-end GPT-6 Luna and Claude Haiku 5.5, besides astatine considerably little cost.
This doesn’t mean that those of america who wage for $100/month subscriptions volition abruptly get the equivalent for 5 bucks. Instead, these are usage-based metrics based connected API usage.
Also: 4 ways Anthropic’s caller Claude Sheets bushed utilizing a Google Sheets connector
Microsoft home-grown AI models could trim its reliance connected OpenAI and Anthropic
Throughout Satya Nadella’s keynote, Microsoft showed slides referencing a wide scope of instauration models that are intended to enactment with Copilot and different Microsoft offerings. Microsoft famously has a multi-billion-dollar concern successful OpenAI dating backmost to 2023, and has built galore of its products and services astir OpenAI intelligence property.
But now, Redmond is benchmarking its internally built models against Anthropic and OpenAI offerings, and besides touting that it’s integrating these caller models arsenic optional choices for GitHub Copilot, Microsoft 365, and Fabric.
Also: 75% of developers I surveyed similar Claude Code implicit Codex
One happening that Nadella made wide successful today’s keynote is however cardinal AI is becoming successful everything Microsoft does. Nearly each of those AI operations antecedently progressive the institution paying immoderate token fees backmost to the frontier exemplary providers oregon customers doing the aforesaid thing.
But with high-performing models of its own, Microsoft whitethorn go substantially little babelike connected extracurricular exemplary providers, particularly for the unthinkable measurement of reasonably basal and regular AI requests that hap crossed the company’s immense merchandise line.
Microsoft turned its coding exemplary into an AI idiosyncratic for Office
If you deliberation backmost implicit the past of agentic enactment models similar ChatGPT Work and Claude Cowork, you whitethorn callback that they originated arsenic agentic coding tools. As users adopted the coding harnesses successful Claude Code and OpenAI Codex to off-label productivity applications, the AI companies noticed and packaged work-centric agentic tools that were much non-programmer-friendly.
Also: I fto ChatGPT Work and Claude Cowork escaped connected my files – lone 1 made maine nervous
There’s a batch of similarity betwixt a coding cause and a enactment agent, including knowing a analyzable environment, utilizing tools to alteration things, and producing reliable results. Coding turns retired to beryllium a precise bully situation for processing and investigating these capabilities, peculiarly due to the fact that codification output is highly measurable. Coding agents are effectual grooming grounds for non-coding enactment agents.
This becomes truly absorbing successful today’s announcement. Along with Code 2 and Code 2 Flash, Microsoft announced MAI Work Flash, which it describes arsenic having been derived from Code 2 Flash. Work Flash tin assistance make formulas successful Excel, enactment with documents and instructions successful Word, way decisions and extract utile accusation from Teams, and more.
Microsoft’s biggest AI exemplary lawsuit mightiness beryllium Microsoft itself
Think astir the level of quality these tasks are demanding from the AI models. Not that much. As agelong arsenic the models are bully enough, they’ll beryllium bully enough.
If Microsoft tin trim its expenses and usage reliable models astatine a overmuch little price, the institution itself volition beryllium capable to prevention rather a spot of wealth crossed its full bundle economy, portion astatine the aforesaid clip maintaining a deeper level of vertical integration and, therefore, control.
Let’s tally immoderate hypothetical numbers. Let’s accidental a institution is spending $0.01 per AI request, which, with these caller models, tin driblet down to $0.001 per request.
Also: Want to effort ChatGPT Wallet? Read OpenAI’s liability disclaimer precise cautiously first
A institution with a regular inference outgo of a cardinal dollars (a institution similar Microsoft, for example) could driblet that regular outgo to astir $100,000. Add that up crossed a year, and the alteration successful models could prevention $325 cardinal annually.
A cardinal here, a cardinal there. Pretty soon, you’re talking existent money.
There’s nary nonstop grounds that Microsoft is trying to nutrient replacement models for the frontier providers’ offerings. But it is an absorbing and compelling communicative if we look astatine Microsoft processing its ain engines for the bundle it sells.
Microsoft’s benchmark claims are impressive, but important questions remain
Microsoft says the numbers it’s reporting astatine are utilizing a 50% introductory discount. That’s wherefore I measured against $0.001 per petition alternatively of $0.0005 per petition above.
Benchmark enactment could beryllium cherry-picking. The numbers are impressive. But arsenic we each know, determination are lies, damned lies, and benchmark results. I’m reasonably definite Microsoft isn’t padding these numbers excessively much, due to the fact that it benefits from the caller models much than adjacent its ain customers, but still. It’s an unfastened issue.
The institution besides described intelligibly that these caller models are not designed to grip the astir challenging agentic and probe workloads. You’re astir apt not going to usage them to lick a decades-old mathematics challenge. But you’re precise apt to usage them to reformat your PowerPoint.
Speaking of reformatting your PowerPoint, Microsoft besides emphasized that these caller models did unusually good connected visualization assignments. That’s a precise bully acceptable for not lone UI plan challenges but besides graphics and descent production.
You tin travel my day-to-day task updates connected societal media. Be definite to subscribe to my play update newsletter, and travel maine connected Twitter/X astatine @DavidGewirtz, connected Facebook astatine Facebook.com/DavidGewirtz, connected Instagram astatine Instagram.com/DavidGewirtz, connected Bluesky astatine @DavidGewirtz.com, and connected YouTube astatine YouTube.com/DavidGewirtzTV.
David Gewirtz
Senior Contributing Editor
Computer subject prof turned AI innovator David Gewirtz has spent implicit 30 years driving advancements astatine the forefront of artificial quality and is the recipient of the Sigma Xi Research Award successful Engineering. He wrote a pioneering investigation that helped signifier the aboriginal speech astir AI ethics. He was besides a pioneer successful the commercializing of AI products, including AI languages and knowledge-based systems. He is the decorator of the Al Editor, an experimental artificial quality motor for parsing, classifying, and identifying quality stories. David is simply a subordinate of the Association for the Advancement of Artificial Intelligence. He presently serves connected the Artificial Intelligence Threat and Mitigation Cross-Sector Council (AI CSC) for InfraGard, a concern betwixt the FBI and manufacture leaders for the extortion of U.S. Critical Infrastructure. He is the writer of Where Have All the Emails Gone?, arsenic good arsenic How To Save Jobs and The Flexible Enterprise. See afloat bio

51 minutes ago
3








English (US) ·