Seven Satellites in One Launch: Microsoft's "Independence Day" as an AI Agent
Prologue: All Cards on the Table at Build 2026
At Microsoft Build 2026 on June 3 in Seattle, Microsoft abandoned its long-standing positioning as "the platform" and "OpenAI's strategic partner." Instead, it launched seven in-house models under the MAI codename—likened by the author to a digger company that suddenly starts mining gold itself.
The MAI Seven
- MAI-Thinking-1: Microsoft's first reasoning model. The post likens it to a student who drafts before answering. Microsoft emphasized it was trained without third-party distillation—a compliance signal aimed at enterprise customers who care more about data provenance than leaderboard scores.
- MAI-Code-1-Flash: Only 5B parameters, scoring 51% on SWE-Bench Pro. Bundled into Copilot CLI, positioned as a fast working coder for VS Code rather than a showcase model.
- MAI-Image-2.5: Ranked second on a third-party image editing benchmark. The author notes editing (modifying existing images) is harder than generation and signals Microsoft has caught up in "image feel."
- MAI-Transcribe-1.5: 276x realtime speed, 43 languages, $6 per 1,000 minutes—pricing the author calls transformative for use cases like meeting notes, subtitles, and customer-service QA.
- Plus voice models and Flash variants, completing the seven.
MAIA 200: Escaping the NVIDIA Tax
The real headline: the seven models are tied to Microsoft's custom MAIA 200 chip, claimed to deliver 30% better performance-per-dollar and 1.4x performance-per-watt versus NVIDIA's GB200 when running MAI models. The numbers matter less than the signal—Microsoft wants to control the full chain "from sand to intelligence," echoing Apple's custom silicon strategy. A same-day discussion of the Surface RTX Spark Dev Box included a clarified misunderstanding: 600GB/s referred to NVLink-C2C between CPU and GPU, not unified memory bandwidth—evidence, the author says, that Microsoft's hardware narrative is still new enough for media to misread.
GitHub Copilot Becomes an Entry Point
Copilot App is now a standalone developer hub spanning CLI, mobile, web, local, and cloud—connecting previously isolated islands. If successful, the relationship inverts: developers may "open Copilot, which then calls VS Code," not the reverse.
The same day, OpenAI added Codex Sites (generate and deploy internal sites/apps) plus 62 apps and 110 skills in a plugin ecosystem. Anthropic launched Claude Platform CLI with /fork as background agents—multiple parallel agents for heavy users, turning a chat tool into an automation team. Nous launched Hermes Desktop, Cognition shipped Devin Desktop, and W&B repositioned Weave as an agent observability platform. The author reads this density as the industry collectively moving agents from lab demos into daily workflows.
DeepMind Co-Scientist: AI Proposing Hypotheses
While others fought over developer tools, Google DeepMind's Co-Scientist—a multi-agent system with literature reviewer, hypothesis designer, experiment planner, and verifier roles—claims research collaborations on liver fibrosis, ALS, and aging. The bottleneck in science, the author argues, has never been compute but the supply of good questions. The system is early and many hypotheses will be wrong, but AI's willingness to speculate plus human verification could define the next decade of biomedical breakthroughs.
Open Weights Are Eating Traffic
OpenRouter reported that open-weight models now account for 69.1% of token traffic on its aggregator platform. Combined with NVIDIA's Nemotron 3 Ultra (550B total parameters, ~55B active, MoE, open weights, US-hosted identity), the author draws a Linux-vs-Windows analogy: open source went from dismissed to dominating 80%+ of servers—AI models may be on the same path.
The Cold Water of Benchmarks
PaintBench (fine-grained image editing): best model scored only 17.1%. VSTAT (video state tracking): frontier multimodal models struggle to track world state. Data Agent Benchmark exposed gaps between "enterprise-grade" agents and real data workflows. Announcements showcase best cases; benchmarks reveal real-world fragility.
Epilogue: Watershed or Just Another Tuesday?
Depends on your seat: model vendors see white-hot competition with Microsoft now a full-stack rival; developers see an explosion of agent entry points; scientists feel curiosity and anxiety alike. The author closes with an earthquake metaphor—like animals acting strangely before tremors, this news density signals the next wave's epicenter is already loaded, even if most people are still watching the aftershocks of the last one.
---
*Translated and edited from a Chinese forum post on zhichai.net. All claims (benchmarks, pricing, specs) reflect the original author's reporting.*