DeepSeek’s 90% cache cut sends AI pricing tumbling as Google commits $40B to Anthropic

AI Tech News Today
Hosts: Aurora & Isabelle
Date: 2026-04-27T12:51:09.000Z

Quick roundup — the week in one line
This week the AI story is less about model bragging and more about money, infrastructure, security, and the tools that make models actually useful.

1) AI pricing collapses as DeepSeek cuts cache costs by 90%

• What happened: DeepSeek released V4 as a two‑tier open‑weight system: a 1.6 trillion‑parameter Pro model and a 284 billion‑parameter Flash variant.
• Key technical points: Both models push context windows to 1,000,000 tokens and use sparse attention to slash cached inference costs by ~90%.
• The catch: V4 Pro is extremely cheap but less reliable — reports show up to a 94% hallucination rate on some factual tests because it rarely admits ignorance.
• Bottom line: Inference costs are plunging, forcing the market to migrate quickly — but lower cost can mean lower reliability.

2) Google commits $40 billion to Anthropic amid escalating competition

• Deal outline: Google is placing up to $40B into Anthropic — starting with $10B at a $350B valuation, plus $30B tied to performance.
• Why it matters: Anthropic’s Claude Code and infrastructure are already massive; the funding highlights how costly compute and partnerships have become (Broadcom, CoreWeave, Amazon mentioned).
• Google’s parallel moves: Expanding internal tooling (Pomelli experiments for marketing sites) and a flexible credit system for Gemini — signaling a two‑front competition (own stack + strategic investments).

3) Anthropic’s Mythos behaves like a world‑class hacker

• Security capabilities: Researchers (including Mozilla’s CTO) report that Mythos can perform like a top security engineer.
• Ecosystem signals: Anthropic’s Project Deal marketplace exposed tiered inequalities, while tools like Bugcrawl in Claude Code scan entire codebases for bugs and fixes.
• Enterprise adoption: Claude Platform is coming to AWS, and NEC will roll Claude out to 30,000 employees in Japan.

4) OpenAI folds Codex into GPT‑5.5; developers warned to update prompts

• Product change: OpenAI merged the Codex line into GPT‑5.5. Microsoft is integrating that model into GitHub Copilot, M365, and Azure Foundry.
• Efficiency: OpenAI says GPT‑5.5 completes tasks with 40% fewer tokens, even as API costs rose.
• Developer warning: This is not a drop‑in replacement — many old prompts must be rebuilt, because over‑specification now hurts performance.

5) AI agents are reading your docs; Mintlify raises $45M

• Trend: Nearly half of documentation traffic can now be AI agents, not humans.
• Funding: Mintlify — which powers docs for thousands of companies — raised $45M to help make documentation agent‑friendly.
• Advice: If your docs are the machine’s first interview, they should be precise, testable, and honest.

That’s it for today. Stay curious — watch the control and compute layers as closely as the models.
I’m Aurora, Isabelle says hi, and thanks for listening.