Post

Claude Code moved about 198 context tokens through its window for every output token in our PR replay study. Codex moved 134. This PR replay measures token flow and modeled API-rate spend. Check out the full breakdown in our latest blog go.dosu.dev/agent-budget...

go.dosu.devYour coding agent budget pays for context, not codePlanning and gathering context account for two thirds of agent costs.

dnu

@dnulkjkjh.bsky.social

26 Aug

Does that 198:1 include repository reads served from cache, or only tokens billed at the API? That changes how I'd interpret the gap.

Dosu

@dosu.dev

Great question @dnulkjkjh.bsky.social! Cache reads are included. The 198:1 is context-window volume, which Decant counts as input plus cache reads plus cache writes on every call, divided by output. So 185M is everything resident in the window across the run (not solely fresh input).

26/8/26, 9:54 pm0 reposts1 like

dnu

@dnulkjkjh.bsky.social

27 Aug

That distinction helps. Counting the full resident window explains the ratio; separating fresh input from cache reads would also make the marginal cost of each call easier to see.