I think part of the reason that claude has better output is that it takes in more as an input(ex reading more of your code so more tokens used). Whats the token limit with gpt5?
You are viewing a single comment's thread from:
I think part of the reason that claude has better output is that it takes in more as an input(ex reading more of your code so more tokens used). Whats the token limit with gpt5?
For reference, for the above denser job mentioned, it took around 4.5 million tokens total for it to get done with Codex with no sign of running out of weekly limit (I counted the cycle of compacted context of 228k tokens over 4 weeks). No clear idea how much it took Claude, but it took almost 20ish days. Claude did it faster (less iterations from my side). My friend whose claude I was using, said I burned 8 million tokens, though I am sure he was using it as well. I almost always ran out of the daily limit of 5 hours with Claude, but not with codex. Even with 4 projects coding parallely did I come to 17% one time.
Now for further reference, Codex has a "fast" mode which I have yet to enable, which basically allows higher expenditure of tokens for faster speed.
In terms of IDE, I found Codex easier to use and less likely to loop itself, which Claude did sometimes. Claude required more proactive statements in the markdown files than Codex, while Codex required more reactive statements.
That said, I prefer codex, but if this is only for claude max 20x then I will take it. I don't care what I code with as long as I have an effective tool. Both get the job done.
Also I have spent 3 months making my own tools which work very well with codex. I maintain my own context through a db and record everything. I have my own framing tools, and a concierge that communicates with the agent to ensure no deviation and so on.