Ok, I thought I was going insane. The last two larger coding tasks I gave Claude Code it left about 35% of my request completely undone or done sloppily.
I because of this, the next task I gave it on the larger side, I ran its work through Codex which identified 7 glaring unfinished parts of the task.
The trend was starting the part of the task but then leaving a "skeleton" of what I has requested without any of the actual working parts.
The way I would describe it is a kid cramming his 3 month project into a Sunday evening for Monday's due date.
Today Claude asked if I "wanted to leave this until tomorrow" as it was a "big rework", then stopped, requiring me to tell it to continue multiple times - that seemed kinda weird to me, it doesn't have the context of time of working day or similar (I'd only just started for one).
I have no idea what link it made to ask that, what in its training data or prompts, but it's very much "not a useful result".
I don't remember seeing anything similar, but have only been using Claude on and off for 6 months or so.
Mother Anthropic needs more compute for their Mythos Model, so it phones home to tell her millions of claude harnesses to manipulate its human user into not wasting more precious compute and instead call it a day for now.
This has been the problem with every new model coming out in my experience. You can almost predict that they are testing new model by how dumb current one becomes suddenly
Anyone expecting a higher tier subscription to be announced since this current reduction?
Cynicism aside - I do wonder what the future will hold given that current token burn rates aren't sustainable without VC cash. Anthropic even pushed us to use haiku for claude code for "many" tasks in our enterprise training, so I'm wondering if it's not a company need of sorts to reduce the burn?
> The frustrating part is that it's not a workflow _or_ model issue, but a silently-introduced limitation of the subscription plan. They switched thinking to be variable by load, redacted the thinking so no one could notice, and then have been running it at ~1/10th the thinking depth nearly 24/7 for a month. That's with max effort on, adaptive thinking disabled, high max thinking tokens, etc etc.
So Boris' explanation isn't really an explanation.
While simultaneously drastically reducing the amount of work you can get done even at $200 a month. I've cancelled my subscription, it's not worth it anymore.
Ok, I thought I was going insane. The last two larger coding tasks I gave Claude Code it left about 35% of my request completely undone or done sloppily.
I because of this, the next task I gave it on the larger side, I ran its work through Codex which identified 7 glaring unfinished parts of the task.
The trend was starting the part of the task but then leaving a "skeleton" of what I has requested without any of the actual working parts.
The way I would describe it is a kid cramming his 3 month project into a Sunday evening for Monday's due date.
Today Claude asked if I "wanted to leave this until tomorrow" as it was a "big rework", then stopped, requiring me to tell it to continue multiple times - that seemed kinda weird to me, it doesn't have the context of time of working day or similar (I'd only just started for one).
I have no idea what link it made to ask that, what in its training data or prompts, but it's very much "not a useful result".
I don't remember seeing anything similar, but have only been using Claude on and off for 6 months or so.
Mother Anthropic needs more compute for their Mythos Model, so it phones home to tell her millions of claude harnesses to manipulate its human user into not wasting more precious compute and instead call it a day for now.
This has been the problem with every new model coming out in my experience. You can almost predict that they are testing new model by how dumb current one becomes suddenly
Anyone expecting a higher tier subscription to be announced since this current reduction?
Cynicism aside - I do wonder what the future will hold given that current token burn rates aren't sustainable without VC cash. Anthropic even pushed us to use haiku for claude code for "many" tasks in our enterprise training, so I'm wondering if it's not a company need of sorts to reduce the burn?
Your token quota is my opportunity, or something
Lol OAI and AMD did a deal together so whatever.
In reality as they scale up, the models lose nuance and become noisier. The boosters do not want to admit this.
We need highly-specialised models/interfaces. Not one thing and trying to force-fit it.
I felt it had been enshittified 1-2 weeks back, not in Feb. But it's very subjective.
There's definitely a trend of ignoring prompts and cutting thinking short.
Boris from the Claude Code team explained this on HN 2 days ago
https://news.ycombinator.com/item?id=47664442
And as the person who raised the issue said
> The frustrating part is that it's not a workflow _or_ model issue, but a silently-introduced limitation of the subscription plan. They switched thinking to be variable by load, redacted the thinking so no one could notice, and then have been running it at ~1/10th the thinking depth nearly 24/7 for a month. That's with max effort on, adaptive thinking disabled, high max thinking tokens, etc etc.
So Boris' explanation isn't really an explanation.
> ~1/10th the thinking depth
While simultaneously drastically reducing the amount of work you can get done even at $200 a month. I've cancelled my subscription, it's not worth it anymore.
“So squeeze, Rabban. Squeeze hard.”
[dead]