Feedback: Significant regression in software design and contextual awareness between versions 1.0.24 and 1.0.61

For users experiencing the regressions in recent versions, it is possible to temporarily force the use of the older, more reliable version (1.0.24) or any other version you comfort with.

npx @anthropic-ai/claude-code@1.0.24
npx @anthropic-ai/claude-code@1.0.24 -v
npx @anthropic-ai/claude-code@1.0.24 [your_command_and_arguments_here]

or

npm uninstall @anthropic-ai/claude-code
npm install @anthropic-ai/claude-code@1.0.24 -g
claude config set -g autoUpdaterStatus disabled
claude

This is can't be long term solution because of:

1. Risk of Deprecation: Package managers like npm allow maintainers to unpublish or deprecate older package versions at any time, often for security or maintenance reasons. Once version 1.0.24 is removed from the registry, this workaround will immediately fail, and the agent will become unusable via this method.
2. No Support or Updates: By locking to an old version, you will not receive any new features, performance improvements, or critical security patches released in newer versions of the CLI.
3. Potential API Incompatibility: Future updates to the backend Anthropic platform or APIs could break compatibility with the older 1.0.24 agent at any moment, causing it to stop working without warning.

Great post i'm also seeing this and was wondering what was happening

Yes me and my team have also been wondering what is going on, since we suddenly got much worse results and claude seems to be a lot less smart, which is really quite critical as we start to rely more on it and invest time to build custom tooling and enhancements!

@PiotrMocan , @fwends @ajbmachon , which alternatives are you guys researching and testing, i'm also looking to switch and cancel after my subscription end

@PiotrMocan , @fwends @ajbmachon , which alternatives are you guys researching and testing, i'm also looking to switch and cancel after my subscription end

I'm not looking at switching just yet I will give them a chance to fix this, they are prolly still the best out there even tho this is clearly an issue.

which alternatives are you guys researching and testing, i'm also looking to switch and cancel after my subscription end

@JohnDotOwl If older version works and give me same experience as month ago I'll keep with claude code, if not qwen 3 coder pretty good model, some providers are generous and gives same amount of usage for reasonable price

Thanks for version suggestion. Downgrade didn't help against overall slowness. CC "thinks" 600+ seconds and not even consume a token :( I don't know what is happening under the hood.

I am also not looking to switch however, if this problem persists I am looking at using claude code router with Kimi K2 instead. I really love the product and would want to stick with it. The max subscription is still worthwhile for me even with the reduced rates that we are getting now, but the drop in intelligence and instruction following is something I cannot deal with permanently.

I wonder if we are silently getting routed to older models, while Anthropic is struggeling with server load, prior to the new datacenters going up. see here [Older Model Names Issue](https://github.com/anthropics/claude-code/issues/4489 ).

The problem is actually building on this product for production makes me very scared now and worst case, I must consider gemini cli or its fork qwen code and adapt them to my usecases.

Learning from my experience with cursor , it's almost required to have a fallback that works. Most company wouldn't have transparency. E.g. the reddit post is now talking about a weekly limit https://www.reddit.com/r/ClaudeAI/comments/1ma3s58/weekly_limits_are_coming/

I'm playing around with Qwen cli (Gemini fork) with qwen coder api , seems to be the only alternative that is up to standard.

I am also not looking to switch however, if this problem persists I am looking at using claude code router with Kimi K2 instead. I really love the product and would want to stick with it. The max subscription is still worthwhile for me even with the reduced rates that we are getting now, but the drop in intelligence and instruction following is something I cannot deal with permanently. I wonder if we are silently getting routed to older models, while Anthropic is struggeling with server load, prior to the new datacenters going up. see here Older Model Names Issue. The problem is actually building on this product for production makes me very scared now and worst case, I must consider gemini cli or its fork qwen code and adapt them to my usecases.

I've tested Kimi K2 + claude code, with newer versions of claude code doesn't matter what model you will use, they compress context x5-x6 times

> I am also not looking to switch however, if this problem persists I am looking at using claude code router with Kimi K2 instead. I really love the product and would want to stick with it. The max subscription is still worthwhile for me even with the reduced rates that we are getting now, but the drop in intelligence and instruction following is something I cannot deal with permanently. > I wonder if we are silently getting routed to older models, while Anthropic is struggeling with server load, prior to the new datacenters going up. see here Older Model Names Issue. > The problem is actually building on this product for production makes me very scared now and worst case, I must consider gemini cli or its fork qwen code and adapt them to my usecases. I've tested Kimi K2 + claude code, with newer versions of claude code doesn't matter what model you will use, they compress context x5-x6 times

Ah ok thanks for the heads up. Its great to know, that its due to the application not the models.

Do any of you have a specific prompt or test case that demonstrates the change in output that you're observing?

The system prompt has changed between versions, but the underlying model is the same. It's possible but unlikely that you are seeing different results, but the only way to tell would be to compare it on the same code base with the same prompts multiple times. Otherwise it's likely that the cause is differences in the prompts and code.

Note that I'm not excluding the possibility that you're right, but I would very much like to be able to test the claims.

So, if you can provide a repo + commit + the prompts used, or even a simple reproducing example unrelated to your work, that would be very helpful for analyzing and tracking the changes!

Do any of you have a specific prompt or test case that demonstrates the change in output that you're observing? The system prompt has changed between versions, but the underlying model is the same. It's _possible_ but unlikely that you are seeing different results, but the only way to tell would be to compare it on the same code base with the same prompts multiple times. Otherwise it's likely that the cause is differences in the prompts and code. Note that I'm not excluding the possibility that you're right, but I would very much like to be able to test the claims. So, if you can provide a repo + commit + the prompts used, or even a simple reproducing example unrelated to your work, that would be very helpful for analyzing and tracking the changes!

Yes, I do, you absolutely right that result of prompt execution is non determined, but I see patterns in responses, ant I tell you more, newer versions compresse context, thats easy to check, just run roo code and select cloud code as provider, you will see how much context consumes your promts with different claude code version, in my cases overal context consumed by exact same prompt with 1.0.61 reduced by 5.6 times, and that maybe the cause of such different results

We've investigated this and haven't seen a widespread, consistent regression in behavior. While prompt changes may cause behavior to vary on a case-by-case basis. In general, the most helpful thing here would be to file new bugs with complete session info for sessions where Claude is not doing what you expect it to.

tested new 1.0.89, I have to admit results are better, though still provides half working solutions

<img width="1783" height="1096" alt="Image" src="https://github.com/user-attachments/assets/5609c707-544b-4b0d-9c86-c827c904ef13" />

older version
<img width="1794" height="1163" alt="Image" src="https://github.com/user-attachments/assets/689161d8-fd37-4846-b9be-90002e9c9e8e" />

here the repo with prompt https://github.com/PiotrMocan/testcc

context usage

<img width="1338" height="398" alt="Image" src="https://github.com/user-attachments/assets/7869c523-e436-43d0-a490-d2ebeb0d3436" />

<img width="645" height="251" alt="Image" src="https://github.com/user-attachments/assets/cd7b0b67-1831-495a-9654-5e40a28f046a" />

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.