Today in AI — 5 April 2026
Today's top AI news — curated links and commentary on the stories that matter for product builders.
OpenAI spent the week losing executives, killing products, and retiring beloved models. Everyone else spent it building.
OpenAI's rough patch
Three senior departures in a single day while the company prepares for an IPO at $852B. Add the Sora shutdown that blindsided Disney and GPT-4o's final retirement, and you have a company managing more crises than product launches.
- OpenAI loses three top executives at once as IPO looms — TechCrunch
- Altman says he 'felt terrible' telling Disney about Sora shutdown as $1B deal dies — Variety
- GPT-4o is fully retired — the model 800,000 people didn't want to lose — OpenAI
Anthropic plays offence
A $400M biotech acquisition, a published multi-agent harness for autonomous coding, and a new PAC for AI policy in Washington. Three fronts in one week.
- Anthropic acquires biotech startup Coefficient Bio for $400 million — TechCrunch
- Anthropic ships a three-agent harness for autonomous full-stack development — Anthropic
- Anthropic files to create AnthroPAC as AI midterm spending tops $300 million — TechCrunch
AI with almost no humans
A two-person telehealth startup tracking toward $1.8B revenue. A state chatbot renewing psychiatric prescriptions. AI is replacing human labour faster than regulation can follow.
- One man, two employees, and $1.8 billion: the AI-built company is here — PYMNTS
- Utah gives an AI chatbot the power to renew psychiatric prescriptions — Gizmodo
Breaking the hardware bottleneck
DeepSeek is porting V4 entirely to Huawei chips while PrismML's 1-bit models run at 131 tokens/second on a laptop. One route is geopolitical necessity; the other is pure engineering. Both reduce dependence on Nvidia.
- DeepSeek V4 will run entirely on Huawei chips, breaking from Nvidia — Reuters
- PrismML launches the world's first commercially viable 1-bit LLMs — The Register
Agents, safety, and scale
Agent infrastructure is maturing fast. Moonbounce raised $12M to moderate AI companions, Claw Code hit 100K GitHub stars after Anthropic's source leak, and the DOE is putting frontier models in the hands of 1,000 scientists across nine national labs.
- Moonbounce raises $12M to build content moderation for the AI companion era — TechCrunch
- Claw Code becomes fastest repo in GitHub history to hit 100K stars — Claw Code
- OpenAI deepens Department of Energy partnership with 1,000-scientist AI jam session — OpenAI
The gap between companies building AI and companies built by AI is closing fast. If your product still assumes humans in every loop, the economics are shifting under you.