Today in AI — 3 April 2026

The AI industry stopped debating which model is best and started competing over who owns the most surface area. In a single day, three companies shipped new model families, OpenAI pushed further into media and advertising, and the security industry raced to lock down an agentic layer that barely exists yet.

Three model families in a day

Microsoft, Google, and Alibaba each released new models within hours of each other. Microsoft's MAI series is its most direct move into model development. Google shipped Gemma 4 under Apache 2.0, stripping away earlier commercial restrictions. Alibaba's Qwen 3.6-Plus matches frontier benchmarks at $0.29 per million input tokens. More choice, lower costs, less lock-in.

• Microsoft takes on AI rivals with three new foundational models — TechCrunch
• Google DeepMind launches Gemma 4, its most capable open models under Apache 2.0 — Google Blog
• Alibaba unveils Qwen 3.6-Plus with 1M-token context to accelerate agentic AI — Caixin Global

OpenAI builds the conglomerate

OpenAI bought a media company, restructured developer pricing, and crossed $100M in annualised ad revenue. Platform-building, not model development.

• OpenAI acquires TBPN, the buzzy founder-led business talk show — TechCrunch
• OpenAI rolls out Codex-only seats with pay-as-you-go pricing, cuts ChatGPT Business to $20 — OpenAI
• ChatGPT ads hit $100M annualized revenue as self-serve access opens this month — OpenAI

Securing the agentic layer

AI agents create identity and trust problems that existing security tooling cannot handle. Cisco's DefenseClaw sandboxes agent tools before deployment. Linx raised $50M to map human, non-human, and agentic identities across enterprises. Tennessee banned AI from posing as mental health professionals.

• Cisco launches DefenseClaw and Zero Trust for AI agents at RSA Conference — Cisco Newsroom
• Linx Security raises $50M Series B as identity becomes the top AI security problem — SecurityWeek
• Tennessee bans AI from posing as mental health professionals as states accelerate chatbot regulation — Transparency Coalition

Capital and infrastructure

Q1's $297B in venture funding shattered records, with foundational AI startups accounting for $178B alone. Microsoft's $10B Japan commitment shows capital flowing into physical infrastructure: data centres, GPU hosting, workforce training.

• Q1 2026 shatters venture funding records at $297B as AI captures the majority of all VC — Crunchbase News
• Microsoft invests $10B in Japan to build AI infrastructure and train 1 million AI professionals — Let's Data Science

Platform and tooling moves

Google now lets you import ChatGPT and Claude conversation history into Gemini. Anthropic is going vertical with a leaked biology workspace called Claude Operon. And Hugging Face shipped TRL v1.0, giving post-training a stable production framework.

• Google lets you import ChatGPT and Claude conversation history into Gemini — Google Blog
• Anthropic's Claude Operon leaked — a dedicated biology research workspace inside Claude — Geeky Gadgets
• Hugging Face releases TRL v1.0, turning LLM fine-tuning into a production framework — Hugging Face

The model layer is commoditising, the platform layer is where power concentrates, and the security layer is still wide open.

---