The Brief

New briefs daily around 7 AM Eastern

Google ships Gemini 4 Hydrogen at $1.50/$7.50, above Argon on AA index

Hydrogen replaces Gemini 3.8 Flash and lifts Terminal-Bench 4.1 from 18.2% to 57%, and it is now the free-tier model. It costs 25% less than Carbon and Argon today and 75% less after their Dec 1 doubling, yet 15x Haiku 5.5's input price.

• 01Read

  Jensen Huang leaves Nvidia CEO role; Colette Kress named interim CEO

  SemiAnalysis tracks over 1.5GW of restricted Nvidia chips that reached China and expects RASA-style export rules to advance without Huang. Many US startups build on Chinese open models trained on Blackwell GPUs, so tighter rules reach that supply.

  Dylan Patel→
• 02Read

  Mercor's OWE-Bench: Fable 5.2 leads 26 long engineering tasks at 54.6%

  Fable 5.2 also spends less per attempt than GPT-6 Astra, $105 against $235, for ten more points. Across 624 trajectories, most failures came from agents testing on easier inputs than the hidden cases, then shipping work that broke at real scale.

  Mercor research blog→
• 03Try

  Claude Code 2.1.323 adds Code Mode, calling tools by writing JavaScript

  Plugins can now declare a Claude Code version range in plugin.json, and clients outside it skip them, so a plugin update no longer breaks older installs. Long hooks in claude -p and SDK sessions send a 30-second heartbeat, so hosts stop ending those turns as hung.

  claude-code v2.1.323→claude-agent-sdk-typescript v0.3.323→
• 04Try

  AI SDK 7.0.171 stops malformed UI messages from exhausting server memory

  Any app that validates client UI messages was exposed: a small request with many malformed parts made validation report every schema for every part. The release also keeps Azure and Vertex credentials from reaching third-party file hosts during downloads.

  github-actions[bot]→
• 05Read

  Vercel lays out a three-tier test pyramid for agents and their tools

  Request logs only show that a call succeeded, so an agent that orders a hot latte for an iced request looks healthy. Grading repeated end-to-end runs measures how often the task is done right, which a single manual test cannot show because outputs vary.

  George Fahmy→