Jev cuts Opus 5.5 thinking tokens by 98.5% by picking reasoning levels
Across six coding tasks, Jev chose the reasoning level per step while Opus 5.5 wrote the code, cutting thinking tokens by 98.5%, running about 2.4x faster and passing 12/12 hidden checks.
i think i found where Opus 5.5 wastes its thinking Jev + Claude Opus 5.5 passed the same coding checks with 68x fewer thinking tokens the setup caught my attention because Opus still does the work. Jev decides how hard it needs to think before the next step six coding tasks. two runs per task. the published results: → Opus 5.5 at max → 64,899 thinking tokens → 913 seconds → Jev + Opus 5.5 → 954 thinking tokens → 387 seconds → both setups → 12/12 runs passed the hidden checks 98.5% fewer thinking tokens. about 2.4x faster at API rates, the total went from $3.76 to $1.70. 55% less for the same passing results one task made the difference especially clear: 54,397 thinking tokens → 551 $2.00 → $0.32 remove that task and the cost still fell 21% here's the part I'd use in my own agent: > Jev reads the current task and recent results > picks the reasoning level for the next step > Opus reads the code and makes the changes > tests check the work > Jev chooses again as the task moves forward the reasoning level changes without throwing away the prompt cache the comparison matters: these were small tasks against fixed max effort. fixed high was already close to the hybrid 526 seconds saved across this benchmark. $2.06 less at API rates keep Claude on the code → let Jev choose the reasoning level → set up the pair with this guide ↓
This post quotes another post: x.com/bimbacrypto/status/2103843888532230575