Today in AI — 28 April 2026

Google committed up to $40 billion to Anthropic. The same week, DeepSeek cut its API prices by 90%. The capital flowing into AI and the cost of running it are moving in opposite directions, and the gap tells you where value is actually accumulating.

Capital picks sides

Google's $10 billion commitment to Anthropic, with $30 billion more tied to performance targets, is less an investment than a hedge against losing the next platform. Cohere merged with Aleph Alpha in a sovereign AI deal backed by two governments, while David Silver raised a record $1.1 billion seed to build AI that learns without human data.

• Google to invest up to $40 billion in Anthropic in cash and compute — CNBC
• Cohere acquires Aleph Alpha in $20B sovereign AI deal backed by two governments — TechCrunch
• DeepMind veteran David Silver raises record $1.1B seed for Ineffable Intelligence — TechCrunch

The price floor drops

DeepSeek cut cache-hit prices to one-tenth and offered 75% off V4-Pro, while its open-source V4 models match frontier performance with 27% of the compute. OpenAI's GPT-5.5 reclaimed some benchmark leads, but the direction is clear: intelligence gets cheaper faster than anyone building on it expected.

• DeepSeek slashes V4 API prices by up to 90% in aggressive price war — Bloomberg
• Three reasons why DeepSeek's V4 open-source model matters — MIT Technology Review
• OpenAI releases GPT-5.5 with improved coding and agentic capabilities — CNBC

Hardware for the next layer

OpenAI is reportedly building a phone where agents replace apps entirely, targeting 300–400 million annual units. Google split its TPU line into dedicated training and inference chips, a design choice that only makes sense if you believe agent workloads will dwarf batch processing.

• OpenAI reportedly building AI smartphone where agents replace apps — TechCrunch
• Google unveils eighth-generation TPU chips designed for the agentic era — Google Blog

Trust still trails capability

Sam Altman apologised to Tumbler Ridge after OpenAI failed to alert police about a ChatGPT user who carried out a mass shooting that killed eight people, Anthropic published a postmortem on a Claude Code regression, and Simon Willison traced the death of the AGI clause that would have let OpenAI cut Microsoft off from its most advanced models. The systems are acting in the world; the governance is still catching up.

• OpenAI CEO apologizes to Tumbler Ridge for not alerting police about mass shooter — TechCrunch
• Anthropic publishes postmortem on Claude Code quality regression — Anthropic
• Simon Willison traces the death of the OpenAI-Microsoft AGI clause — Simon Willison

Developer tooling

Anthropic is testing Bugcrawl, a Claude Code feature that runs 10 parallel agents to scan repositories for bugs and propose fixes. The tool that writes your code is learning to audit it too.

• Anthropic tests Bugcrawl, a parallel bug-hunting agent for Claude Code — Testing Catalog

The money says control the platform; the market says intelligence is a commodity. If you're building products, bet on what you do with the models, not on locking in a provider.

---