Claude thinks before it answers, and one setting decides how hard: the effort. On Claude Opus 5.5 it goes from low to max, thinking cannot be switched off, and every thinking token is billed as output, at $20 per million tokens on Anthropic's public rates. On a subscription seat the same thinking uses up the plan's rate limits faster.
The levels are far apart. TokenSave had Opus 5.5 build the same small browser game at each one (their write-up). Medium, the model's default, took 52 seconds and cost $0.56. High cost $0.75 and extra high $1.36. Max ran for 22 minutes and cost $5.25, nine times medium, for 25 times the output tokens. All five games worked. That is one run of one small task, but it agrees with Anthropic's own advice: start at medium, go to high where the quality falls short, and keep max for problems that need it.
On a team, effort is each developer's own choice in /model, made one session at a time. Someone who leaves max on spends several times what their colleagues do on the same work, and until last week nothing on the Logs screen showed which level a request had used.
Set a ceiling for each model maker
Settings → Models now opens on an Effort limits table. It has one row for each model maker with a model switched on: Anthropic, OpenAI, Moonshot and so on. Each row shows the effort levels the maker's switched-on models take, and a Max effort select: No limit, Low, Medium, High or Extra high. The select offers only the levels that would lower something those models take, and it saves as soon as you pick.
The ceiling belongs to the maker, not to the provider that serves its models. One OpenRouter key can serve Claude, GPT and Kimi side by side, and a ceiling on Claude should not hold back the other two. A ceiling on Anthropic follows Claude everywhere it runs: a Claude subscription, your own Anthropic key, Vertex or OpenRouter.
Once a ceiling is set, Selan lowers any request to that maker's models that asks for more, to the highest level the model takes at or under the ceiling. A request that asks for less keeps its level, and one that names no effort runs at the model's own default.
The pickers follow it
Codex offers only the levels at or under the ceiling, and a model whose default was above it starts at the highest level left. Claude Code can only hide extra high and max, so under a Medium ceiling it still offers High, and Selan lowers those requests to Medium. Any other client reading /v1/models sees the same capped levels.
See the level each request ran at
Logs now show the effort beside the model, and the request's details list it too. When a ceiling lowered a request, Logs show the level it actually ran at. A request that left the choice to the model shows none.
Pick a level and stay on it
Each model in Claude Code's picker starts at its own default effort, and a level a developer picks in /model stays theirs on every later launch. Anthropic's documentation warns that changing the effort in the middle of a conversation invalidates its prompt cache, so the next turn pays to write the whole conversation again. Choose the level when the session starts, not halfway through it.
The ceiling is set per company and per model maker, by an owner. What an owner sets lists it with the rest of the settings.