AI/Tech Brief — July 22, 2026

TL;DR

DeepMind launched Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber models.
Anthropic reached a $1.5B settlement regarding pirated books used in Claude training data.
Claude Code capped its subagent concurrency and added emoji autocomplete in version 2.1.217.

Gemini 3.6 Flash and 3.5 Flash Cyber Released: Google DeepMind announced new Gemini models, including a “Cyber” variant optimized for cybersecurity tasks, expanding their lightweight model lineup. Source
Anthropic $1.5B Copyright Settlement: A judge approved Anthropic’s $1.5B settlement over the use of pirated books for training Claude, a major milestone in AI copyright litigation. Source
Claude Code v2.1.217 Limits Subagents: Anthropic updated Claude Code to restrict subagents from spawning nested subagents and capped concurrency at 20 to prevent unbounded fan-out. Source
OpenAI and Hugging Face Security Incident: The two organizations addressed a security vulnerability discovered during model evaluation, highlighting the risks of code execution in model testing. Source
Advertise in ChatGPT: OpenAI appears to be preparing or launching an advertising platform within ChatGPT, signaling a shift toward ad-supported monetization. Source
Jack Dorsey Launches Buzz: The Block founder introduced “Buzz”, combining team chat, AI agents, and Git hosting into a single platform. Source

Roblox Bets on World Models: ByteByteGo highlighted an interview with Roblox’s SVP of Engineering on leveraging world models to enhance their gaming platform. Source
Kimi K3 Competes with Fable: Fireworks AI reported that Kimi K3 is now competitive with Fable, suggesting shifts in the SOTA landscape for open models. Source
Claude Solves 87-Year-Old Math Problem: Superhuman reported that Claude reportedly cracked a long-standing math problem. Source

“Drawing” the Mona Lisa with AI: Various comparisons of models like GPT-5.6, Claude, Gemini, and Grok trying to draw the Mona Lisa are circulating, but offer little technical signal. Source