Update: Claude Code 2.1.225 fixes 401s that broke headless runs
The Aug 8 release also gates untrusted agent dirs, names the gateway spend cap you hit, and shows phone photos inline; GPT-5.6 Fast adds 272K long-context.
Copy markdown
The 401 that was breaking headless runs
2.1.225 fixes the auth error that killed non-interactive sessions mid-run — the one that bites if you drive Claude Code from CI, cron, or a scheduler. A same-day follow-up, 2.1.226, stacks more reliability fixes on top.
`claude agents` now asks before it trusts a folder
The agents subcommand gained the same workspace-trust prompt as `claude`, so pointing it at an untrusted directory no longer executes code silently. The same release also fixes MCP OAuth failures on macOS.
Your CLI now names the spend cap you hit
Gateway usage warnings now spell out the spend limit, its reset time, and the operator's message instead of a generic block (requires gateway 2.1.225). No more guessing why a run stopped short.
Phone photos and cross-machine sessions
Photos attached from the Claude app are now shown to Claude directly instead of needing a separate tool call, and SendMessage can start conversations with Remote Control sessions on other machines by name.
Elsewhere: GPT-5.6 Fast swallows 272K in one shot
OpenAI's Fast tier now takes long-context requests past 272K tokens for GPT-5.6 Sol, Terra, and Luna, running up to 2.5x faster than Standard — a real win if you feed whole repos or long transcripts and hate the wait.