Daily AI Builder Brief - April 22, 2026

Executive Summary

Claude Code 2.1.117 shipped today: agent frontmatter mcpServers now load on main-thread --agent sessions, and /model selections persist across restarts even when a project pins a different model.
Default effort on Opus 4.6 and Sonnet 4.6 is now high for Pro and Max subscribers (was medium), so your existing prompts will think longer without you changing anything.
Forked subagents are behind CLAUDE_CODE_FORK_SUBAGENT=1 on external builds; concurrent connect to local and claude.ai MCP servers is now the default.
Bloomberg reports an unauthorized group accessed Anthropic's unreleased Claude Mythos Preview via a third-party vendor environment on the same day the model was announced to a 40-company consortium.
Anthropic is A/B testing removal of Claude Code from the $20 Pro plan with 2% of prosumer signups; plan-pinning logic in your CI may need a look.
/resume now offers to summarize stale, large sessions before re-reading, and WebFetch no longer hangs on very large HTML pages.

Claude Code

2.1.117: agent MCP loading, forked subagents, high-effort default

Agents with mcpServers: in frontmatter now load those servers when invoked as the main-thread agent via --agent, closing a gap from 2.1.116. /model selections persist across restarts even against a project pin. Forked subagents can be enabled on external builds via CLAUDE_CODE_FORK_SUBAGENT=1. Default effort on Opus 4.6 and Sonnet 4.6 is now high for Pro and Max (was medium). On native builds, Glob and Grep tools are replaced by embedded bfs and ugrep behind the Bash tool. Fixes: WebFetch hangs on huge HTML, inflated /context percentages on Opus 4.7, Plain-CLI OAuth sessions dying on token expiry. Claude Code changelog

Anthropic

Claude Mythos Preview: unauthorized access via third-party vendor

Bloomberg reported April 21 that a small group accessed Mythos Preview, Anthropic's offensive-cyber frontier model, through a third-party vendor environment on the day of announcement. The group guessed the URL based on Anthropic's naming conventions, with help from an employee at a vendor. Anthropic says there's no evidence of impact to core systems. Builders with vendor-hosted inference or red-team tooling should revisit URL-predictability and vendor-side access logging. TechCrunch, red.anthropic.com

Claude Code pricing test: 2% of Pro signups see it removed

Anthropic is testing removal of Claude Code from the $20 Pro plan with ~2% of new prosumer signups. Nothing is final, but if your onboarding docs or CI assume Claude Code works on Pro, add a fallback check. The Register

MCP Ecosystem

Concurrent MCP connect is now default in Claude Code

Startup is faster when both local and claude.ai MCP servers are configured: concurrent connect is now the default path. Pair this with 2.1.117's agent-frontmatter loading, and per-agent MCP surfaces get a lot cleaner. Claude Code changelog

Broader AI

Nothing today.

What This Means For Builders

Upgrade to Claude Code 2.1.117 and audit any agents that rely on mcpServers: in frontmatter; they now work on main-thread --agent runs.
Your Opus 4.6 / Sonnet 4.6 prompts will think longer by default on Pro/Max. Budget for it, or set effort explicitly.
If you distribute internal tooling built on Claude Code + Pro, add a plan-detection fallback now; the pricing test signals the shift.
Treat vendor-hosted model URLs as predictable by default. Add access logging and reject speculative paths.