Are you reviewing the code or just let fly whatever it produces? If you’re reviewing and not running parallel sessions, I can’t imagine how you manage to hit that limit.
DaAnalyst
DaAnalyst
Weird. Wouldn’t say it differs much from the way I use it, except for maybe the time I spend on reviewing the code (which is substantial).
0
manhvu
I no longer pay for Claude — it’s too expensive. GLM from Z.ai is good enough for me and much cheaper than Claude/GPT/Gemini. To save even more money, I usually add “Don’t explain, just update the source and summarize” to the rules/system prompt. That way, I can use LLMs all day for only $10–20.
3
asianfilm
I’m not the person you are replying to, but just to note that I got close to my 20x limit for the first time last week when writing code generation in Elm: in other words Elm code that writes Elm code.
As I got to that last 10% of code generation, I had to develop a bunch of techniques to stop Claude from taking shortcuts, with one Claude Code instance commissioning and signing off on another’s work.
I’ve never been warned about being close to my limit when writing Elixir. But maybe I’m controlling context well by using private Hex packages to control the versioning of the layers of code in my stack.
Generally I agree that those limits are never hit for me in the past 12 months in normal usage where I’m reviewing Claude’s work, since I’m in a conversation and getting new ideas as we pair program.
1
Vidar
I have two long running projects of which at least one pretty much always works. Data intensive so they take time but are not always heavy on tokens. Certain prompts can take 6 to 12 hours even on a rather decent computer. These are the paying ones so I do follow them very carefully, and also do test runs up ahead before the main branch. Those optimizations and explorations following many branches can take a lot of tokens.
As for the rest that varies. Some are play projects to learn while some are connected to Archdo or the skills. (I’ve put out about 8-9 skills or so publically, but we have closer to 40 internally and there is always something to improve on).
Then there are potential future projects which I probe around, and of course old pet projects which suddenly became easy to do. 
Long term my aim is indeed to get good code straight out from LLMs, but not quite there yet. It is getting closer though.
Edit: And yes, I do have parallel sessions. Some are always on standby in their own terminals, while up to three are actively running.
2
arcanemachine
Less than $100. And I use Claude Code on a daily basis.
Considering the quality difference between Anthropic and OpenAI models, I would rather build on a platform that allows me to use their service how I see fit (i.e. in other harnesses) without having to pay the massively-more-expensive API pricing (currently ~70x more after Anthropic increased usage limits after their deal with xAI). Not by constraining me to using their subpar TUI that breaks every other week (the skills autocompletion UI broke for me in a new release, within the past week).
Conversationally, I love Opus (still on 4.6). I just wish I could use it with a decent harness.
Dollar for dollar, I get much more mileage out of an OpenAI subscription than I do by using Claude.
1
DaAnalyst
DaAnalyst
what is it that you don’t like about Claude Code’s harness? Me, I’m totally happy with it.
0
dimitarvp
Seeing that you are working on a GUI framework, where the LLM has no closed (or at least tight enough) feedback loop, your stance is very understandable – and normal, I’ve heard it from multiple other people doing GUI work.
When you can give Claude a good test harness and a near-perfect requirements list / acceptance criteria, it does mostly well.
Sadly the more autonomous sessions leave Claude vulnerable to “this broke but it’s fine because…” or “this does not look good but it’s here before us so I am leaving it” or “I don’t like this but it’s out of scope” (for problems that almost human dev would immediately claim are in scope).
It’s not perfect. I am still hesitant to try GPT 5.5 but likely will; having two LLMs and one of them at least doing adversarial / critical reviews is by itself quite valuable.
3
johns10davenport
I remain super bullish on Claude Code overall.
Partly because I don’t have to think about it. It just works, and I’m almost always under token budget.
The level of value I get out of the tool is huge, I do:
- Massive amounts of engineering work
- All my content writing for multiple websites
- Significant amounts of marketing (both strategy and execution)
- SEO audits and improvements
- Personal finance setup
And lots more. Of course, this isn’t intrinsic to Claude. I could probably use Pi + API tokens and potentially save some money, but I’m paying for it to work and not to think about it basically.
2