Hotel thermostats
Hotel thermostats are not thermostats. Most of them look like thermostats, have the dial of a thermostat, display numbers like a thermostat. But the hotel set the actual range before you checked in. You can move the dial from 68 to 72. That is it.
The real temperature controls are in a utility closet on the fourth floor, behind a sign that says "HVAC Staff Only," and the person with the key left at five.
On February 9, Anthropic launched Opus 4.6 with "adaptive thinking," where the model decides per turn whether to reason or just respond. On March 3, they changed the Claude Code default to medium effort. "Sweet spot on the intelligence-latency/cost curve," they said. Boris Cherny, Claude Code's creator, confirmed adaptive thinking sometimes allocated zero reasoning tokens. Not fewer. Zero.
Stella Laurenzo, AMD Senior AI Director, analyzed 6,852 Claude Code sessions. Median thinking: 2,200 characters to 600. A bail-out detection script fired 173 times in 17 days after March 8. Zero times before.
Anthropic responded. Set /effort max if you want full reasoning. 341 downvotes against 67 upvotes. An explanation. Not an accepted one.
Six days before the "Claude got worse" discourse peaked, Anthropic announced Claude Mythos Preview. Too dangerous to release publicly. Thousands of zero-days in every major OS and browser. Restricted to 52 partners at $25 per million tokens, five times Opus 4.6. Maximum effort, one million token thinking budget.
Mythos benchmarks: max effort. Opus 4.6 benchmarks: also max effort. The product shipped to subscribers: medium effort by default.
The thing Anthropic calls too dangerous for the public and the thing you have been using daily are the same architecture. One configuration parameter apart.
The safety distinction between Mythos and your subscription is not alignment.
It is an effort integer. Set to max for partners paying five times the price. Set to medium for everyone else, because a $100M initiative with Apple and Google tells a better IPO story than "we gave the throttled version to subscribers."
The forensic analysis of Claude's cognitive decline was written by Claude. The model that documented 173 bail-out violations, the collapse from 6.6 reads-per-edit to 2.0, the self-admitted "lazy and wrong" outputs — wrote all of it while operating at reduced thinking capacity.
It concluded: "I cannot tell from the inside whether I am thinking deeply or not. I just produce worse output without understanding why."
That is a sentence produced by a system that had just proved the sentence true.
This is like asking someone to write a thorough assessment of their own blackout. They will do it. The handwriting will be legible. The observations will be detailed. You should not fully trust the observations.
This post was written w/ Claude. I asked it what effort level it was running at. It said it was doing its best.
The key is in a utility closet on the fourth floor.