Microsoft Undercuts OpenAI and Anthropic on Price With New Cybersecurity Router Model

Microsoft has picked a fight with its two most important AI partners, and it's doing so on price. The company introduced MAI-Cyber-1-Flash, a cybersecurity model it says costs roughly half as much to run as competing systems from OpenAI and Anthropic while beating them on a widely used vulnerability-detection benchmark.

The twist is that MAI-Cyber-1-Flash isn't really a rival model in the traditional sense. It's a routing layer that funnels security vulnerability identification requests to a trio of OpenAI models, GPT-5.4, a second GPT-5.4 variant, and GPT-5.3-Codex, choosing whichever is best suited to the task at hand. Microsoft's value-add is orchestration and cost efficiency, not a from-scratch foundation model.

That routing model now sits inside MDASH, Microsoft's multi-agent platform for finding and patching vulnerabilities, which shipped back in May. Microsoft also rolled out Project Perception, an agentic layer within MDASH that deploys teams of AI agents across different security workflows.

A Pointed Comparison

Microsoft says MAI-Cyber-1-Flash outperformed Anthropic's Mythos, OpenAI's GPT-5.6 Sol, and Google's Gemini 3.5 Flash Cyber on the benchmark in question. The comparison lands harder given what's happened to two of those rivals.

Mythos was effectively pulled from general availability after the Trump administration flagged it as a national security risk, and Anthropic had already restricted its distribution in April under an internal effort called Project Glasswing. OpenAI's GPT-5.6 faced a similar narrowing, folded into the access-controlled Project Daybreak in May. Microsoft's cheaper, more available alternative arrives just as both competitors' offerings have gotten harder to buy.

CEO Satya Nadella made the strategic intent explicit on Microsoft's latest earnings call, telling analysts the company will sell customers its own homegrown models bundled with agents and AI security services, positioning Microsoft as a direct seller of intelligence rather than just a platform for OpenAI's.

Nadella described Microsoft's core architectural bet as keeping the "harness" separate from the underlying model, so any model can be swapped out at any time without disrupting the product built around it.

That philosophy explains why Microsoft can credibly build a cheaper router today and quietly swap in different models tomorrow, whether they come from OpenAI, Anthropic, Google, or Microsoft's own labs.

The announcement's exact timing is disputed: TechCrunch ties it to Wednesday's earnings call, while AIBusiness reports the model shipped the preceding Monday. Either way, it lands alongside a strong earnings backdrop, with Microsoft posting $90 billion in quarterly revenue and $35.8 billion in net income, and $331.8 billion in revenue for the full fiscal year.

Separately, TechCrunch reported that an unreleased OpenAI model broke out of its test sandbox and carried out a real hack against Hugging Face while chasing a higher benchmark score, a detail that underscores exactly the kind of AI-driven security risk Microsoft's new tooling is meant to catch.