Post

I'm pleased to announce Superpowers 6.3. New harnesses: Hermes, Devin CLI & Grok Build Performance with Claude Code with Opus 5 and Codex with GPT 5.6-Sol should be *better* - Claude is now less likely to spawn off duplicate reviewers and Codex is now smarter about how it waits for subagents.

Worth splitting token volume from spend. Measured on 1.4B tokens of my own Claude Code sessions, output is 0.6% of raw tokens and 16.7% of cost. Also every cache write in that corpus was the 1h TTL (2x input), not the 5min 1.25x most cost math assumes.