Microsoft Undercuts Its Own Suppliers on Price

The slide carried two numbers and no adjectives. Microsoft said its new coding model reaches 96% of the capability of Anthropic’s Claude Opus 5.5 and sells for 5% of the price.

The comparison was the sharpest moment of a keynote otherwise devoted to product packaging. At its Ignite conference on Nov. 17, Microsoft released seven models built by its in-house MAI team, and priced every one of them against a competitor it also resells.

MAI Code 2, the coding model, reaches 98% of the performance of GPT-6.1 Sol across four enterprise coding benchmarks, Microsoft said, at 40% of the price. Against Opus 5.5 the claim is 96% of capability at 5%. A second model, MAI Cyber 2 Flash, matched GPT-6 Astra on the CyberGym end-to-end security test at under 2% of the cost per task, according to the company.

Those figures are Microsoft’s own, measured on benchmarks Microsoft chose, and they have not been independently reproduced. That is normal for a launch and also the reason to hold them loosely. Benchmark parity in coding tasks has historically translated unevenly into performance on the messy repositories that enterprises actually maintain.

The pricing is the part that requires no verification. A company that owns a model can price it at marginal cost, which for inference is the electricity and the depreciation on the chips. A company reselling someone else’s model pays a per-token fee that includes that supplier’s margin. The gap between the two is where Microsoft is now competing.

The security model is the one to watch for a different reason. CyberGym measures whether a system can find and exploit vulnerabilities across an end-to-end target, and a model that performs at a fraction of a competitor’s cost changes the economics of both attack and defense. Microsoft sells to governments and to banks, two buyers that care about cost per test less than they care about who holds the data.

The strategic logic has been visible since 2025, when Microsoft began building models in-house while continuing to sell OpenAI’s. The company has described its approach as a router: send each request to whichever model does the job at the lowest cost. Building its own models changes the router’s default setting.

For OpenAI the arrangement has always carried an awkward clause. Microsoft is its largest investor, its largest distribution channel and, as of this week, a competitor selling a substitute at a fifth of the price on the same storefront.

Microsoft did not frame the launch that way. Executives described the models as options for customers who want to keep sensitive data inside their own tenancy, and for workloads where the cheapest adequate model is the right answer. Both arguments are commercial rather than technical, and both point at the same buyer: the enterprise information-technology department trying to hold a budget.

The rest of the keynote was about where those models live. Copilot Workspaces puts the assistant into a shared surface where documents, meetings and agents are worked on together, and Microsoft described Copilot as the operating system for work, with Office applications embedded inside it rather than arranged around it.

That is a bundling argument, and bundling is how Microsoft has won before. The company’s advantage in enterprise software has never been the quality of an individual application. It has been the fact that the applications arrive together, priced together, administered together, and that removing one is harder than adding all of them.

Satya Nadella opened the keynote by apologizing. An Outlook outage the previous week had taken email down for users worldwide, and the chief executive addressed it before he addressed anything else.

Reliability is now part of the sales pitch rather than a footnote to it. A model that is 96% as capable and available every hour of the quarter is worth more to a bank than a better one that fails without notice. Microsoft’s own record on that measure has been uneven this year.

The company also announced a partnership with Salesforce, under the name Copilotforce, to connect their respective agent platforms. The two firms have spent two decades as adversaries in customer software, and the agreement suggests that both now see the agent layer as a market large enough to share.

What none of this settles is whether enterprises will move. Multi-model routing is cheap to configure and hard to govern: an organization that sends different tasks to different models has to document which data went where, and regulators in several jurisdictions now ask for exactly that record.

The pricing claims will be tested in procurement cycles rather than in benchmarks. A 5% price against Opus is a bid, and bids get compared. If the capability gap widens on the next generation of models, the discount will have to widen with it.

Microsoft’s shares did not move much on the announcements, which is the market’s usual verdict on a product launch that reframes a strategy rather than changing a forecast. The quarterly report that follows will show whether customers agreed.

Related Posts

  • November 20, 2026
  • 18 views
Sandisk Completes Eight-for-One Split as NAND Outlook Darkens

Eight for one. Sandisk’s stock split took effect on Nov. 19, and the shares began trading on the adjusted basis on Nov. 20, a change that divides the price by…

  • November 20, 2026
  • 16 views
A Lawmaker’s Privacy Suit Puts Model Design on Trial

For three years British courts have sidestepped the question of whether the company that builds an AI model answers for what its users produce with it. A written defence filed…