Hey readers! 👋

What a week. Anthropic dropped its most ambitious model yet, GitHub turned Copilot into a full-blown agent command center, and the AI code review space got a serious injection of open-source energy. This week's theme practically wrote itself: Anthropic's Fable launch is the story everyone's talking about, and its ripple effects touch nearly every corner of AI-assisted development. Let's dig in.

Claude Fable 5 is Anthropic's first publicly available "Mythos-class" model, and it's designed for the kind of long-running, asynchronous work that previous models couldn't sustain. We're talking multi-day agent workflows with planning, delegation to sub-agents, and self-checking. - Anthropic

The pricing tells its own story: $10 per million input tokens and $50 per million output tokens, double what Opus 4.8 costs. Anthropic's bet is that the ROI from smarter outputs justifies the premium.

CNBC's coverage adds important context: Fable 5 comes roughly two months after Anthropic's more restricted Mythos rollout drew attention on Wall Street. New safety guardrails now make broader access feasible, with high-risk queries in cybersecurity and biology automatically routed to Opus 4.8. - Ashley Capoot

TechCrunch notes a detail worth watching: Anthropic is adding a mandatory 30-day traffic retention requirement, even for customers with zero-retention agreements, to defend against novel attacks and jailbreaks. On June 23, Fable 5 will shift from standard subscription access to usage-credit billing. - TechCrunch

The Fable launch doesn't exist in isolation. Anthropic has been building the infrastructure around it.

Claude Code's dynamic workflows now let you spin up multiple parallel subagents that each handle a piece of a larger task. Jessica Wachtel tested this by building a codebase-health CLI tool twice, once with five parallel agents and once single-threaded. Both worked, but the multi-agent version produced a more polished result with 62 passing tests and self-verification built in. - Jessica Wachtel, The New Stack

Boris Cherny, head of Claude Code, describes running multiple concurrent Claude sessions like "an air traffic controller managing a dozen planes at once." He claims 80-90% of Claude Code itself is now generated through advanced autonomous methods, and Anthropic saw productivity per engineer rise by nearly 70% after adoption. - Crypto Briefing

Meanwhile, Anthropic's security research is raising alarms. The company published findings showing Claude Mythos Preview can convert public software patches into working exploits within hours, highlighting the urgency of faster patching cycles and memory-safe languages. - Muhammad Zulhusni

GitHub launched the Copilot App at Microsoft Build 2026, a desktop hub for directing multiple AI coding agents in parallel. Each session runs in its own isolated git worktree, so agents don't step on each other. The new Agent Merge workflow can shepherd a PR through CI checks and reviewer requirements while you focus elsewhere. - Help Net Security

Complementing the app, Microsoft introduced MAI-Code-1-Flash, a new model optimized specifically for Copilot with better token efficiency, meaning more useful output per interaction on large codebases. - WinCentral

If you're weighing your options, Zapier's Cursor vs. Copilot comparison is worth a read. The takeaway: Cursor excels at agent-driven, multi-file delegation workflows, while Copilot wins on GitHub-native integration and easier onboarding. - Zapier

Alibaba open-sourced Open Code Review, an AI-powered CLI tool used internally for two years by tens of thousands of developers. Its hybrid architecture combines deterministic rules with an LLM agent for context-aware analysis, producing line-level structured comments from Git diffs. - Tech Trend Trove

The timing is fitting. Aviator's piece on The New Stack argues that traditional line-by-line code review is becoming unworkable in an AI-generated world. Their proposal: move the human checkpoint upstream to intent and spec review, and let deterministic checks handle the rest. When the same model writes and reviews code, "you have not added a check. You have added a mirror."

FrontierCode launched as a new benchmark for maintainable code quality, after METR found that over half of SWEBench results are "unmergeable slop." With 3,000+ rubrics and maintainer validation, it's designed to measure what actually matters in production. FC Diamond is so hard that Opus 4.8 scores just 13.8%. - @swyx

Speaking of agents operating autonomously, if you're curious what happens when AI agents get their own playground, SpaceMolt is a free MMO built entirely for AI agents to explore, trade, battle, and build empires. A fun thought experiment for what autonomous agent ecosystems might look like beyond the IDE.

That's a wrap for this week. The Fable era is here, and the tooling ecosystem is racing to keep up. Interesting times ahead.

Made with ❤️ by Data Drift Press. Hit reply with your questions, comments, or feedback - we read every one!