Today in AI — 2 May 2026

Microsoft's new Copilot doesn't pick a model. It routes between Claude and GPT depending on the task, charging enterprises $99 a month for the privilege of not having to choose. That quiet architectural decision tells you more about where enterprise AI is heading than any benchmark release this quarter.

The single-model era is over

• Microsoft 365 E7 and Agent 365 go live with multi-model Copilot using Claude and GPT side by side — GeekWire
• Harvard adopts Anthropic's Claude and plans to phase out ChatGPT Edu — The Harvard Crimson
• Pentagon signs classified AI deals with seven tech giants, keeps Anthropic frozen out — CNN

The pattern is clear: buyers are treating models like commodities, selecting on capability and policy fit rather than brand loyalty. For builders, your integration layer matters more than your model choice.

Agents that don't wait to be asked

• Writer launches AI agents that act without prompts, challenging Amazon, Microsoft, and Salesforce — VentureBeat
• Alibaba's Metis agent slashes redundant AI tool calls from 98% to 2% — and gets more accurate — VentureBeat

The tension here is real: agents need to do more autonomously, but they also need to do less unnecessarily. Getting that calibration right is the actual engineering problem.

AI pushes into physical spaces

• Google Gemini rolls out to four million GM vehicles in one of the largest AI assistant deployments ever — TechCrunch
• Amazon turns product pages into AI call-in shows with 'Join the Chat' — TechCrunch
• SoftBank targets $100 billion IPO for Roze, an AI robotics venture that hasn't shipped a single robot — CNBC

Google and Amazon are embedding AI into cars and product pages where it can prove value immediately. SoftBank's Roze IPO, for a company that hasn't shipped a single robot, has the unmistakable aroma of 2021 SPAC energy, except the cheques are bigger.

Models get cheaper and smaller

• xAI completes Grok 4.3 API rollout with 40% price cut and one-million-token context — VentureBeat
• Poolside launches Laguna XS.2, an open-source coding model that outperforms Claude Haiku on a single GPU — VentureBeat
• Huawei expects $12 billion AI chip revenue as DeepSeek V4 drives orders away from Nvidia — The Deep Dive

The cost of capable inference is collapsing from every direction simultaneously.

AI meets social

• Shapes exits stealth with $8 million to put AI characters in your group chats — TechCrunch

400K monthly active users averaging four hours a day. That engagement number is striking, and probably worth watching more closely than any enterprise deal announced this week.

When your enterprise platform routes between competing models automatically and your open-source alternative runs on a single GPU, the moat isn't the model anymore. It's the orchestration layer, the data, and the judgement about when to act and when to stay quiet.

---