Code graph tools promise to save tokens and time. In my late-2026 test, only CodeGraph used fewer tokens than plain Claude Code.
10% fewer tokens; 6% lower estimated API cost.
As models get better, these tools start only to get in the way.
https://t.co/HxxhThC05n
I tested Graphify and CodeGraph with Claude Code to see whether codebase indexing saves tokens.
Both used fewer tokens during feature work, but code quality differed. One run per setup.
Watch the side-by-side test: https://t.co/ZdHX3Wqc0S
Today, someone finally bought me a coffee at #BuyMeACoffee and I had to refund it back. Why? Because they only pay out through #Stripe, which doesn't work in Ukraine. They removed other withdrawal methods back in 2024. That is saddening. β
#Superpowers v6 won my retest: 78 active minutes vs 148 for my #Pocock retry. 53% fewer tokens.
Pocock built more, but only 3/8 workers loaded the full implement skill. I missed that while it ran.
One feature, with an imperfect retry.
#Claude Code's weekly boost stays at +50% through Sep 13. ClaudeDevs says it then settles at +25% over the old baseline. I tested three token savers against plain #Opus 5. #Ponytail won: 18.1M vs 46.3M tokens, a 61% cut. Stacking all three saved less.
https://t.co/2LDxk4o255
Matt Pocock's #Wayfinder plans deeply before implementation. Addy Osmani's skills build sooner, then spend heavily on review.
Both runs cost about $87. The time didn't match: 290 vs 108 minutes. Neither replaced opening the app and checking the result.
https://t.co/xl92J6Ulzj
I tested #Beads and #Backlog.md against #Claude Code's native Tasks. All handled the work equally well. Storage was the difference: Backlog.md keeps Markdown, while Beads uses #Dolt. Claude's native task files disappear when the list closes.
https://t.co/WVzyyCUKc4
Back to $100. Opus 5 high for planning, medium for execution. It isn't that bad.
And I had to drop the local reviewers' effort to "medium" too (#Opus 5 and #GPT 5.6 Sol); before, it'd take hours to run and produce "iteration limits exceeded".
Do we even need #fable? π±
@cosmintrica That's insane. I never actually used Fable on xhigh, even on 20x. Guess, given enough tools and files to read, it devours limits like crazy
Classic #SDD is dying. #BMAD, #OpenSpec and #SpecKit charge a ceremony tax the new models don't need. Opus 5's own guide says strip the legacy scaffolding out. Tested BMad's new autonomous loop anyway: six hours, still on story three, nothing shipped. https://t.co/XGeTeVdVd9
This is just wrong.
Not so long ago, a $100 #claude subscription used to be sufficient for anything. Now it's only Friday, and I'm done.
They might be quietly cutting down the usage limits after each so-called 2x promo.
Should I move on with #codex as the daily driver?