The Wire / 2026-10-24
Five reads from the October 13 to 24, 2026 window: Cloudflare finds where its review tokens went, Huntley proposes a different bargain for maintainers, and Willison offers two small tool updates and a game worth playtesting.
-
Cutting our inference bill by 68% by being obsessive about optimization
Cloudflare reports an average cost of $0.77 per AI code review, down from $2.43 in its two most expensive weeks before optimization. That figure assumes high effort for its code-quality reviewer. It credits prompt-cache fixes, fewer wasted calls, and model and effort changes. A rising cache-hit rate had masked growing token use. There was a tradeoff: its evaluation found fewer important performance issues at medium effort. If your reviewer is getting expensive, measure tokens and calls per completed review, then compare findings on a fixed sample before accepting the cheaper setup.
-
a new model for open source: closed to pull requests, open to remix.
Huntley plans to disable pull requests on code released to GitHub while keeping discussions open, and proposes that users carry adaptations as Nix overlays. The argument is that cheap patches should not create an unlimited claim on a maintainer's time. If agents are filling your contribution queue, state whether you want a patch, a reproducer or an idea, and who will maintain a downstream change.
-
llm-echo 0.5
Willison describes llm-echo as an LLM debugging plugin that repeats its inputs without calling a model; version 0.5, the post says, expands its simulated token accounting to cover tool calls, system prompts and reasoning tokens. Use a dummy model like this to inspect what your integration sends and test usage reporting before spending tokens on a live run.
-
llm-tools-datasette 0.1.1
Willison says LLM 0.33's switch to httpx2 exposed a missing direct httpx dependency in llm-tools-datasette, repaired in 0.1.1. If your Datasette tool broke after that upgrade, this is the release to try. For plugin authors, test installation in a fresh environment so a dependency supplied by the host does not quietly become part of your contract.
-
Battle Drones - a vibe coded RTS with audio-based multiplayer
Willison reports that Battle Drones went from a dull prototype to a game that is fun to play after repeated rounds of asking Claude to improve the fun and playing the results. By the same account, Claude ran over a thousand simulated playtest sessions, while almost none of the code has had a human review. For a low-stakes agent experiment, borrow the feedback loop: keep playable versions and spend your attention trying the result, with a separate code review if the project grows beyond a toy.