One Pot, One Quota: Codex and ChatGPT Are Moving In Together
Left your coding agent running overnight? You might wake up to a ChatGPT that won't think. Here's what December's shared limit could mean for your workflow (and your patience).
The Hook: Imagine leaving your coding agent running, coming back to an empty tank, and finding you can't even chat with Thinking. Heavy Codex users are bracing for exactly that.
The Meat: Starting in December, ChatGPT and Codex will share one allowance, measured in tokens. Instant chat stays outside the pool, but Thinking chats and Codex runs will drink from the same cup. Work mode draws from Codex usage, so it counts toward the pool too. OpenAI says it expects most people won't wait on a reset any more than they do now. The internet is not convinced.
Spotted on X, the worry came fast: as one user put it, a single coding session could make the whole subscription feel unusable until the limit resets. Another says they burn their weekly Codex limit in 3-4 days, because that's when life lets them code. Will the other days of the week be chat-free? One commenter summed up the mood with a fridge: shared quota is two roommates and one fridge, and Codex is the one eating everyone's leftovers at 3am.
What People Want:
- Separate caps: A dedicated limit for conversational reasoning, so one long agent run can't block unrelated chats.
- Visibility: Clear usage meters before anyone hits zero mid-workflow. Several people asked whether Thinking and Codex will drain the pool at the same rate. We haven't seen an answer.
- Spending controls: Reserve a slice of the pool for chat before a background agent goes off on its own.
The Cynical Take: Some see a monetization play. Merge the quotas and heavy users either upgrade or ration the agent. Others fear Codex is heading toward Pro-only territory.
The Practical Take: One commenter says the real drain isn't the long run, it's the transcript growing underneath it. Step 30 re-sends far more than step 3, so the window dies by compounding. Their fix: cap the agent in steps and force a checkpoint instead of watching tokens. Another asked if two Plus subs, one for coding and one for chat, would work (no one confirmed). And one commenter noted that Claude does this too.
The Takeaway: If you run agents in the background, watch for how OpenAI shows the shared meter in December. Until then, a step cap is your friend.