A-TIER tool
you don’t need Netflix with this tech drama.
100k ft — What is this?
you don’t need Netflix with this tech drama.
50k ft — Why care?
you don’t need Netflix with this tech drama. — @nataindata; uses Anthropic, Claude, Claude Code.
10k ft — Key takeaways
(no takeaways yet)
5k ft — Tools & techniques
Anthropic, Claude, Claude Code
(no techniques tagged)
1k ft — Deep-dive / try it
you don’t need Netflix with this tech drama.
Anthropic hid a switch inside Claude Code that silently downgrades you from Opus to Sonnet when their servers hit load. No notification.
You just keep paying for the smarter model and getting the cheaper one.
—
In February, developers started noticing Claude felt off.
Most people blamed themselves. Bad prompts.
—
Then one developer decided to actually look.
They went through 6 months of their own Claude Code session logs.
6,852 JSONL files. 17,871 thinking blocks. 234,760 tool calls.
And they ran the numbers:
Thinking depth dropped 67% by late February. Before that, median thinking was around 2,200 characters per response. By late February it was 720. By March it was 560.
The read-to-edit ratio collapsed. In the good period, Claude r…
Transcript excerpt
That's true, Klotko is 67% dumber than it was, and there is even a hidden switch inside the code. In February, users started to mention that something is wrong, not broken, just it got shallower, it was reading the files but ignoring instructions, taking the laziest fix instead of the right one, and of course most people blamed themselves. But then one developer went through six months of their own session logs and discovered a well thing and posted that. Thinking depth dropped 67% by late February, but it all went from reading a file before touching it from six times to two. An automated script like catching Klotk being lazy, dodging work or asking like should they continue instead of just like doing that, went from zero to this times in 17 days. But what is a tropic reply? Here's the Boris's official comment, which in a nutshell says, your analysis might be flawed, like yes we change that but it's better for most users and if you want the old behavior you can try this. But do you remember what was inside the leaked code? Exactly, explicit hidden switch. If you hit three server errors, you silently get downgraded from opus to sonnet. But you still pay for a smarter model and get a cheaper one without any notification. There are actually more things but I think that's enough for now. Link to the issue is in the comments, you can take a look and what do you think?