Tuesday, June 3, 2026 is one of the most eventful days the AI industry has seen this year — with all the major players choosing to strike in the same week.
Microsoft kicked off Build 2026 with a flurry of significant announcements, headlined by MAI-Thinking-1 — the company's first in-house reasoning model. It's a Mixture-of-Experts (MoE) model with 35 billion active parameters (out of roughly 1 trillion total), a 128K context window, trained from scratch on fully licensed enterprise data — no distillation from existing models.
The model is available in private preview through Microsoft Foundry, and aims to compete with Anthropic's and OpenAI's frontier models at significantly lower inference costs. Microsoft claims that in blind evaluations, independent raters preferred MAI-Thinking-1 over Anthropic's Claude Sonnet 4.6.
Alongside MAI-Thinking-1, Microsoft introduced a full model family: MAI-Code-1 for coding, MAI-Image-2.5 for image generation, MAI-Voice-2 for speech, and MAI-Transcribe-1.5 for transcription. The move signals a strategic shift — Microsoft is no longer relying solely on external model providers.
What this means
The new models will soon be integrated into Microsoft 365 Copilot, Azure, VS Code, and other products. Microsoft also announced Aion 1.0 — small on-device models (up to 14B parameters) for Windows agents — the Surface RTX Spark Dev Box, an NVIDIA-powered workstation delivering 1 petaflop of AI compute, and Majorana 2, a second-generation topological quantum processor with a target of scalable quantum computing by 2029.
Anthropic expands Project Glasswing
On the other side of the arena, Anthropic announced a dramatic expansion of Project Glasswing — the defensive initiative powered by its Mythos Preview model, a frontier model with exceptional vulnerability discovery and exploitation capabilities. Access will expand from roughly 50 organizations to about 200 across more than 15 countries.
Among the new participants: South Korea, which secured official access for government entities and companies including Samsung, SK Hynix, and SK Telecom. Partners also include Microsoft, AWS, Google, Apple, NVIDIA, and CrowdStrike.
The move raises an interesting debate: should a model capable of human-level vulnerability discovery be restricted — or broadly distributed to protect critical infrastructure?
Regulation enters the room
On Monday, President Trump signed an executive order on promoting AI innovation and security. The order directs federal agencies (including Treasury, Defense, Commerce, and Homeland Security) to seek voluntary agreements with leading AI developers to submit their most advanced models for government cybersecurity testing before public release.
This marks a significant directional shift — a voluntary framework establishing a 30-day review window, signaling a move from a "hands-off" approach toward initial oversight steps. However, industry voices express concern that the order could slow the pace against China.
GitHub Copilot — the backlash
In the background of the big announcements, GitHub Copilot switched to usage-based billing (AI Credits) on June 1 — and the backlash is real. Many users report bill increases of 10x to 60x, with some threatening to migrate to competitors like Cursor or Claude Code.
The shift reflects a broader trend: AI pricing is moving from flat-rate subscription to consumption-based models, similar to cloud computing. The question is whether the market is ready for it.
The bottom line
Today illustrates more than ever that AI is no longer in the "demo" phase. Microsoft is launching serious in-house models, Anthropic is deploying a global cybersecurity safety net, the U.S. government is starting to regulate, and developers are confronting the real costs of usage. The industry is growing up — and this week might be the moment everyone realized it.
