Supercharging your max subscription!
retained text
https://preview.redd.it/ersx8iy1t4vh1.png?width=2160&format=png&auto=webp&s=cd15ee41114a2f007d283d534e481a6391b0dc78 https://preview.redd.it/rekllgsps4vh1.png?width=2160&format=png&auto=webp&s=6d98177c272ea5f7d730aebd96f70511bae9909a https://preview.redd.it/7zq6lhsps4vh1.png?width=2160&format=png&auto=webp&s=4b7764458a39805ba3e5132f9f1d44418e730187 https://preview.redd.it/7yrs5hsps4vh1.png?width=2160&format=png&auto=webp&s=e2fdec52d70cc5df180d98c3a048b98d83a8b633 https://preview.redd.it/gkkmlgsps4vh1.png?width=2160&format=png&auto=webp&s=b18ab48f8a041db083e48b94bc0d540de2ddf01d https://preview.redd.it/pa747ntps4vh1.png?width=2160&format=png&auto=webp&s=17b558dc54fc412161fc49f28ce282d54a990815 https://preview.redd.it/srkijgsps4vh1.png?width=2160&format=png&auto=webp&s=293ee54341373d1d7e73d50fb10d1cddb01410f0 Its that time! I have some token to burn so why not share! All visuals have been generated by Claude based on the run transcripts over the past month! We are experimenting with a Claude setup that runs fully on its own. I logged every session for a month to see what one finished task actually costs and wanted to share our setup so that the community can benefit. We have chopped and changed this a ton but are really happy with where we are. This isn't anything I am trying to sell, or for you to download - just things that have worked for us! **How it works:** every x hours a scheduled Claude Code session wakes up with no memory. It reads a status file and a task queue from the repo, then works through up to 4 tasks of roughly equal size. Each task goes planner → builder → reviewer → a fresh second reviewer, and nothing merges without that last check. The run writes down what it did and exits, and the next one picks up from there. **The numbers:** since Sep 12 that’s 181 runs, 2,278 subagents and 463 merged tasks. A finished, reviewed task went from $33 to $6.90 in compute at API prices (−79%) while also subsequently going up in quality. **Opus 5.5 is the star.** Moving the lead session and the builders from Opus 5 at xhigh to Opus 5.5 at medium took a task from $23–33 to $7.45 in one step. Tokens per run halved, and the final reviewer found fewer serious bugs, not more. **For the Max plan,** that’s about $6.7k of API cost in a month on a $200 plan. Each task now needs a fifth of the compute it did at our peak, so the same plan does roughly 5× more work compared to what we got done a month back! **Key details:** * Opus 5.5 is magical, we deprecated Fable completely * One agent per job, with only the tools it needs * Default effort first, raised only where mistakes are costly * Sonnet only for fully spelled-out steps * Short always-loaded instructions; know-how lives in skills * Test before you cut: blind-graded reviews at high caught as many bugs as xhigh for 39% less EDIT - Github link - [https://github.com/priyanshujha/subagent-studio](https://github.com/priyanshujha/subagent-studio)
document fetched 2026-10-14 00:41 · canonical https://reddit.com/r/ClaudeCode/comments/1x4ixl9/supercharging_your_max_subscription