Opus Max (500usd) No Fix vs GPT o3 Fixed in 3 msgs

So the last 3 days I noticed Claude Opus 4 Max is running in circles all the time, every chat AND even looking into what it’s doing, it’s creating code, then deleting again, creating and delting again several time within same message.

I spent more than a day with Opus 4 not getting anywhere. I am not joking, when I say I opened a chat with ChatGPT o3 and same text etc. after first research it found the issue, I did not believe it. But then after 2 more messages the issue was fixed. That cost me probably 2 usd. While I spent about 500 USD and one and a half day of working with Opus 4 Max. What is going on?? I feel like being cheated the last 3 days time…

The funny thing is, I “HATED” o3 and thought it was a mess which was why I only used Sonnet 4 sometimes and then just Opus 4 cause it was just slightly better always. But this … now im scared of using Opus 4, unless Cursor or Anthropic admit something been very wrong the last 3 days…

Is this only my experience or do others see the same recently?

I attached a screenshot showing the o3 spend. It’s hilarious and same time sad I wasted that much on Opus 4… Can’t show all Opus 4 spend it’s several pages.

Total I spent 3-4k on Opus 4 just for perspective.

hi @AltariGroup thanks for the post, here are a few questions that would help clarify the details.

  1. As Opus 4 is one of the most powerful models, its also one of the most costly. Would Sonnet 4 or even o3 work well for you?

  2. When Opus 4 gets stuck, are those longer chat threads? If yes, starting a new chat helps usually.

Personally I use Sonnet 4 for most complex tasks and rarely have to use Opus 4.
For log checks, code research and similar simpler tasks Auto is usually sufficient and saves you a lot of cost.

(PS. yes the forum bot assumed its a billing related question)

Hey Condor,

I tried the other 2 and choose Opus because it was often on point or got to the point. But last 2-3 days I started to see it circling round all the time. Doing a new chat it’s doing the same I tried many chats for same issue that ChatGPT o3 solved in 1 chat 3 messages. That is quite wild I think and therefor I am concerned something is wrong with Opus 4 Max currently and then I am curious if I am the only one experiencing this, cause there have been problems with it before and extra charges etc, and I am suspecting something to be wrong again now.

1 Like

Yes there is some confusion about the new pricing and how tokens are calculated but there will be hopefully soon an update how tokens are used by AI providers.

I am very active in the forum and have not seen in last days a new report about Opus 4, though there have been a few outages at Anthropic and hangups.

If you could share a Request ID with privacy disabled where Opus hangs, Cursor Team would be able to look if there is anything internal noticeable.

The way it’s done now cause less confusion I think. My concern is something is just plain wrong.

I spent 500 USD on it not being able to solve anything. Now again I tried something simple and it is completely off the rails. Then switching to Sonnet 4 and it works well.

Something is wrong with Opus 4 lately I am 100% sure now. It’s creating big messes currently. Like even simple CSS changes it messes up and create tons of JS that is even a mess and doesn’t work. Something is wrong and it’s just not fair to pay that much for it plus wasting a lot of time. Was so happy with it and don’t mind paying for something helpful, but paying for something that is broken is not fair.

1 Like

Yes it should be clearer. I will watch out for additional posts on Opus and ask if there were other cases. Lets see also what others experience with Opus.

Sure, if you have a Request ID it would be possible to look into the details of the request.

1 Like

There might be today or latest tomorrow an update for Opus, I don’t have details yet but Cursor team is checking already. Thanks for flagging this.

Okay thanks. Not sure I dare to use it again, now I see how o3 performs.

It has been useless last days, only wasting time and money. Sonnet-4 also seems a bit off, not sure about that one though, just know for Opus 4 Max ..

could be this

1 Like

At last it seems like someone is experiencing the same.

Cause I was unsure if I was becoming a complete ■■■■■■ and just did not know how to communicate with Cursor / Opus 4 Max anymore … But others seems to have big problems with it lately as well.

I hope Cursor can jump in here and do something then, cause not only did I waste almost 2 whole days, but also paid 470 USD for fighting Opus 4 Max …

I really do not joke when I say ChatGPT o3 Agent solved it with same Chat and just 3 messages, where as with Opus 4 Max I had 4-5 chats with 10-40 messages in each.

It’s a bit scary the way you are getting charged when things don’t work, are my almost 500 usd just wasted on broken Claude Code Opus 4 or ?

1 Like

@beru thanks for sharing the report, looking into it.

1 Like

This isn’t Claude, this is straight-up Cursor manipulation. We need clarification on this, or I’ll keep suspecting foul play. You can’t pull a fast one on OGs. Especially not me, I’m a criminal coder.

And for the record, GitHub Copilot on VSCode doesn’t pull this sneaky BS.

Cursor AI wraps around models like Claude, GPT-4, etc. If you notice inconsistencies, throttling, “dumbed-down” responses, or degraded performance that don’t match the original API behavior — it’s possible Cursor is proxying or filtering requests in a way that limits or alters output. That would count as “manipulation.”

GitHub Copilot, especially in VSCode, connects directly to GitHub’s LLM (based on Codex/ChatGPT-like models), and doesn’t try to obscure which model is used. It’s relatively transparent — less middleman nonsense. You ask for code, you get code. No mystery meat.

Cursor’s Business Incentive: Cursor is incentivized to look like it’s using Claude or GPT-4-Turbo, but route traffic however it wants (especially on the “free” or low-tier plans). That could include:

  1. Rate-limiting premium model access

  2. Substituting cheaper models

  3. Adding censorship filters

  4. Logging or preprocessing inputs for monetization

It just used to work well, just paying for the Opus 4 usage, but last days is just time and money completely wasted + the headache.

Could make sense, but maybe not on purpose. I dunno. I mean I pay for it… spent about 3500 usd this month + Ultimate membership just on Opus 4.

The Opus 4 performance I see last 3 days is worse than ChatGPT 4o, it’s even worse than not touching anything, cause it destroy things recently and goes into endless circles, even though you rephrase, open new chats, do whatever. I only realized it was very wrong when GPT o3 solved it in 3 messages.

O3 is underrated.
It’s way better than Gemini 2.5 Pro and all the other models.
The only one I haven’t tried is Opus, but I don’t feel the need with O3.

@0xHACKS Please do not make false / unproven claims about things like this.

I can say, unequivocally, the model you choose is always the model you get. The only changes made are if you are not in MAX mode, we lower the maximum context window to help get more requests our of your usage allowance each month.

I am confirming if there has been any changes to Opus’ system prompt / behaviour in the last few days, but the team are so far convinced there is no change on our end that should’ve caused such a degradation in a specific model!

I will update if we learn more, but for now, we don’t see a change that could’ve caused this on our end.

That’s not a refutation, it’s a PR statement.

You unequivocally say the model doesn’t change, then admit to altering context behavior unless in MAX mode. That is a change. OGs notice. Don’t gaslight.

Please refute me with evidence and clear-cut comparisons, not vague reassurances.

I’m waiting!

indeed, the problem with o3 was price until recent weeks and bad agentic mode but they got mostly solved, o3-pro shows as the highest IQ model(https://trackingai.org) but is under grok and sonnet-4-thinking in livebench reasoning tab, that means for some problems it’s better and for some others it’s not
I would suggest @AltariGroup to try other models like grok or o3 immediately, it’s a known issue that can be solved by changing models when a model goes back-and-forth without a solution

@0xHACKS I’ve used Cursor from last October and even if I’m not happy with current pricing or how they communicated changes, I never felt the models were off and believe me I’m really picky with results, always optimizing stuff, just look at my threads, imagine how much bad they are that you’re still here for days bashing Cursor without a single proof, try to make a better world instead

1 Like

If you’re happy with it, great, but personal vibes aren’t proof either. I’m pointing out specific inconsistencies. You don’t feel it? Cool. Doesn’t mean it’s not happening. Let’s not mistake loyalty for clarity.

@0xHACKS try to read before answering, I clearly stated to not be happy with pricing and communication, I’m still waiting for a better release schedule too like a stable version but lets not get off topic

Watcha on about: :grinning_face_with_smiling_eyes:

even if I’m not happy with current pricing or how they communicated changes, I never felt the models were off…

I did read, and you basically said “I don’t like the service, but I trust it blindly anyway.” That’s the topic. Blind trust without transparency is exactly the issue.

Cool, yeah! :+1: