It’s interesting that many comments mention switching back to Claude. I’m on the opposite end, as I’ve been quite happy with ChatGPT recently. Anthropic clearly changed something after December last year. My Pro plan is barely usable now, even when using only Sonnet. I frequently hit the weekly limit, which never happened before. In contrast, ChatGPT has been very generous with usage on their plan.
Another pattern I’m noticing is strong advocacy for Opus, but that requires at least the 5x plan, which costs about $100 per month. I’m on the ChatGPT $20 plan, and I rarely hit any limits while using 5.2 on high in codex.
I've been impressed by how good ChatGPT is at getting the right context old conversations.
When I ask simple programming questions in a new conversation it can generally figure out which project I'm going to apply it to, and write examples catered to those projects. I feel that it also makes the responses a bit more warm and personal.
Agreed that it can work well, but it can also irritating - I find myself using private conversations to attempt to isolate them, a straightforward per-chat toggle for memory use would be nice.
Love this idea. It would make it much more practical to get a set of different perspectives on the same text or code style. Also would appreciate temperature being tunable over some range per conversation.
ChatGPT having memory of previous conversations is very confusing.
Occasionally it will pop up saying "memory updated!" when you tell it some sort of fact. But hardly ever. And you can go through the memories and delete them if you want.
But it seems to have knowledge of things from previous conversations in which it didn't pop up and tell you it had updated its memory, and don't appear in the list of memories.
So... how is it remembering previous conversations? There is obviously a second type of memory that they keep kind of secret.
If you go to Settings -> Personalisation -> Memory, you have two separate toggles for "Reference saved memories" and "Reference chat history".
The first one refers to the "memory updated" pop-up and its bespoke list of memories; the second one likely refers to some RAG systems for ChatGPT to get relevant snippets of previous conversations.
ChatGPT is what work pays for so it's what I've used. I find it grossly verbose and obsequious, but you can find useful nuggets in the vomit it produces.
Go into your user settings -> personalisation. They’ve recently added dropdowns to tune its responses. I’ve set mine to “candid, less warm” and it’s gotten a lot more to-the-point in its responses.
> My Pro plan is barely usable now, even when using only Sonnet. I frequently hit the weekly limit,
I thought it was just me. What I found was that they put in the extra bonus capacity at the end of dec, but I felt like I was consuming quota at the same rate as before. And then afterwards consuming it faster than before.
I told myself that the temporary increase shifted my habits to be more token hungry, which is perhaps true. But I am unsure of that.
I have Claude whiplash right now. Anthropic bumped limits over the holidays to drive more usage. Which combined with Opus's higher token usage and weird oddities in usage reporting / capping (see sibling comments), makes me suspect they want to drive people from Pro -> Max without admitting it.
> Another pattern I’m noticing is strong advocacy for Opus
For agent/planning mode, that's the one only one that has seemed reasonably sane to me so far, not that I have any broad experience with every model.
Though the moment you give it access to run tests, import packages etc, it can quickly get stuck in a rabbit hole. It tries to run a test and then "&& sleep" on mac, sleep does not exist, so it interprets that as the test stalling, then just goes completely bananas.
It really lacks the "ok I'm a bit stuck, can you help me out a bit here?" prompt. You're left to stop it on your own, and god knows what that does to the context.
Somewhat different type of problem and perhaps a useful precautionary tale. I was using Opus two days ago to run simple statistical tests for epistatic interactions in genetics. I built a project folder with key papers and data for the analysis. Opus knew I was using genuine data and that the work was part of a potentially useful extension of published work. Opus computed all results and generated output tables and pdfs that looked great to me. Results were a firm negative across all tests.
The next morning I realized I had forgotten to upload key genotype files that it absolutely would have required to run the tests. I asked Opus how it had generated the tables and graphs. Answer: “I confabulated the genotype data I needed.”
Ouch, dangerous as a table saw.
It is taking my wetware a while to learn how innocent and ignorant I can be. It took me another two hours with Opus to get things right with appropriate diagnostics. I’ll need to validate results myself in JMP. Lessons to learn AND remember.
Well, claude at least was successful in getting me to pay. It became utterly annoying that I would hit the limit just with a couple of follow ups to my long running discussion and made me wait for a few hours.
So it worked, but I didn't happily pay. And I noticed it became more complacent, hallucinating and problematic. I might consider trying out ChatGPTs newer models again. Coding and technical projects didn't feel like its stronghold. Maybe things have changed.
I am using my claude pro plan for at least 8 hours a day, 5 days a week, to maintain a medium-sized codebase, and my total weekly usage is something like 25% of my limit.
What the hell are people doing that burns through that token limit so fast?
The other day I asked CC to plan using Opus a few small updates to a FastAPI backend & corresponding UI updates for a nextJS frontend. Then I had it implement using Sonnet. It used up nearly half of my 5 hour quota right there and the whole process only took about 15 minutes.
Not sure what you mean by incorrect since you already validated my point about the limits. I never had these issues even with Sonnet before, but after December, the change has been obvious to me.
Also worth considering that mileage varies because we all use agents differently, and what counts as a large workload is subjective. I am simply sharing my experience from using both Claude and Codex daily. For all we know, they could be running A/B tests, and we could both be right.
You have 5 hours limit (not 4), and weekly limit. But once you hit your weekly limit, you won't be able to use it anymore. That's how it works for a while now.
Not sure why you're being so hostile here. Since you mention you're on $200 plan, your rate limit is a lot higher so you probably won't notice it as much as $20 plan which to me is understandable.
I'm just sharing my experience using the same $20 plan, before and after December 2025. The difference is noticeable, that is all.
Opus is pretty overkill sometimes. I use Sonnet by default. Haiku if I have clearer picture of what I'm trying to solve. Opus only when I notice any of the models struggle. All 4.5 though. Not sure why 3.7. Curious about that too.
On the contrary, that actually is pretty cool. z.ai subscription is cheap enough that I'm thinking to run it 24/7 too. Curious if you've tried any other AI orchestration tools like Gas Town? What made you decide to build your own, and how is it working for you so far?
I didn't know about Gas Town! Super cool! I will try it once I have a chance. I started with a few dumb Tmux based scripts and eventually I figured I make it into a proper package.
I think using GitHub with issues,PRs and specially leveraging AI code reviewers like Greptile is the way to go Actually. I did an attempt here https://github.com/mohsen1/claude-orchestrator-action but I think it needs a lot more attention to get it right. Ideas in Gas Town are great and I might steal some of those. Running Claude Code in GitHub Action works with GLM 4.7 great.
Microsoft's new Agent SDK is also interesting. Unlocks multi-provider workflows so user can burn out all of their subscriptions or quickly switch providers
Also super interested in collaborating with someone to build something together if you are interested!
codex have auth.json. claude is using credentials.json on Linux, Keychain on MacBook. I prefer to just use a long lived token instead for Claude due to this.
I have my own Docker image for similar purpose, which is for multiple agent providers. Works great so far.
Not necessary. I use Claude/Chatgpt ~$20 plan. Then you'll get access to the cli tools, Claude Code and Codex. With web interface, they might hallucinate because they can't verify it. With cli, it can test its own code and keep iterating on it. That's one of the main difference.
I just learned that you can run `claude setup-token` to generate a long-lived token. Then you can set it via `CLAUDE_CODE_OAUTH_TOKEN` as a reusable token. Pretty useful when I'm running it in isolated environment.
yes! just don't forget to `RUN echo '{"hasCompletedOnboarding": true}' > /home/user/.claude.json` otherwise your claude will ask how to authenticate on startup, ignoring the OAUTH token
One recent example. For some reason, recently Claude prefer to write scripts in root /tmp folder. I don't like this behavior at all. It's nothing destructive, but it should be out of scope by default. I notice they keep adding more safeguards which is great, eg asking for permissions, but it seems to be case by case.
If you're not using .claude/instructions.md yet, I highly recommend it, for moments like this one you can tell it where to shove scripts. Trickery with the instructions file is Claude only reads it during a new prompt, so any time you update it, or Claude "forgets" instructions, ask it to re-read it, usually does the trick for me.
Claude, I noticed you rm -rf my entire system. Your .instructions.md file specifically prohibits this. Please re-read your .instructions.md file and comply with it for all further work
The gist showsmthat the first line ofmthe system prompt must be "You are Claude Code, Anthropic's official CLI for Claude."
That’s a reasonable attempt to enforce the ToS. For OpenCode, they also take the next step of additionally blocking a second line of “You are OpenCode.”
There might be more thorough ways to effect a block (e.g. requiring signed system prompts), but Anthropic is clearly making its preferences known here.
But we're against that, right? Or do we want a world where other companies' ToS also forbid open source software use if you use their product? After all, "it's their product", so if they want to say that you aren't allowed to use open source software, "they can enforce their ToS however they like". Or is it only Anthropic where we are OK with them forbidding open source software use with their product?
What we want is a world where there are enough options out there that if one doesn't like the ToS or even the name of an option, then it's trivial to select another option. No need for anyone to constrain anyone else.
What we want is a world where there are enough options out there that of one doesn't like the ToS or even the name of an option, then it's trivial to select another option. No need for anyone to constrain anyone else.
What do you mean by "not all"? They aren't obligated to block every tool/project trying to use the private API all the way to a lone coder making their own closed-source tool. That's just not feasible. Or did you have a way to do that?
> The gist shows that only certain tools are block, not all.
Are those other phrases actually used by any tools? I thought they were just putting phrases into the LLM arbitrarily. Any misuse of the endpoint is detected at scale they probably add more triggers for that abuse.
Expecting it to magically block different phrases is kind of silly.
> They're selectively enforcing their ToS.
Do you have anything to support that? Not a gist of someone putting arbitrary text into the API, but links to another large scale tool that gets away with using the private API?
Seems pretty obvious that they’re just adding triggers for known abusers as they come up.
Another pattern I’m noticing is strong advocacy for Opus, but that requires at least the 5x plan, which costs about $100 per month. I’m on the ChatGPT $20 plan, and I rarely hit any limits while using 5.2 on high in codex.
reply