Your Codex usage did not evaporate. On ChatGPT plans, Codex runs two limits at the same time: a five-hour window and a weekly cap. Both need room. Empty either one and the CLI stops accepting work, even when the other still shows most of its allowance. That dependency is the part most "usage dropping" threads miss. Waiting five hours can change nothing when the weekly cap is the one that ran out.
Open Codex and run /status. It shows what remains on each limit and when the next reset lands. From there you have three exits: spend the reset, buy headroom, or move the session to per-token billing with an API key. This guide covers how to tell which cap you hit, why the meter runs faster than your diff suggests, and what each exit costs.
Disclosure: Wallaby Token sells API access to Kimi K3, one of the metered options in the last section, and the endpoint in that config block is ours. Everything attributed to OpenAI or to a GitHub thread is quoted from the linked source, fetched October 22, 2026.
The two caps are not alternatives
A user on the Codex tracker put the rule in one line on October 20, 2026 (issue #55480):
If my 5-hour allowance runs out but I have 80% of my weekly allowance remaining, I am blocked. If my weekly allowance runs out but my 5-hour allowance is fully available, I am still blocked.
Two limits, both enforced. The five-hour window absorbs bursts; the weekly cap sets the pace for the week. They are not a menu where one tops up the other. The status line in the CLI labels them primary and secondary and prints the remaining percentage for each. A green five-hour bar means nothing if the weekly number reads 7d 2%.
So ask which clock ran out before asking how to get more. A 40-minute coding session and a five-day week hit different walls.
Which limit did you hit?
The banner tells you which system is talking. These strings come from codex-cli 0.154.0; the current line at the time of writing is 0.163.0, and interface copy moves between releases. Treat the wording as a version-bound sample and the categories as the durable part.
| What you see | What it is | What to do |
|---|---|---|
You've hit your usage limit. |
A plan cap, either the five-hour window or the weekly cap. | Run /status; wait for that reset, or redeem one if you have it. |
You've hit your usage limit. Upgrade to Pro ... purchase more credits |
The same cap, with the paid route attached. | Credits or an upgrade; see the second exit. |
You're out of credits. Your workspace is out of credits. |
Workspace credit balance, not the plan window. | An admin adds credits. |
HTTP 429 with an API key |
Your provider's rate limit, not the ChatGPT plan. | Back off, retry, or rotate the key. |
context window exceeded or compaction notices |
Not a usage cap, but the recovery loop is expensive. | Start a new thread, then read the next section. |
The banner cannot name the window. Only /status shows whether the five-hour number or the weekly number is the one at zero.
One case deserves its own row: a plan cap plus purchased credits. Credits can carry a session past the limit, and issue #48394 (September 26, 2026) asks OpenAI for a confirmation before that happens. Until such a setting ships, the balance can move without a prompt. Check it before you leave a long run unattended.
Why the meter runs faster than your diff
Codex meters the model's reading load: instructions, files, tool output, and the conversation so far. Your typed prompt is a small part of it. A one-line request in a 60-turn thread can cost more than a page of instructions in a fresh one.
Four mechanisms account for most of the gap.
Re-sent context. Every turn re-sends the working set. A long session pays for the same files again and again, and code you did not change still rides along.
Reasoning effort. The CLI warns that a high setting can quickly consume Plus plan rate limits. Effort multiplies output tokens, and output is the expensive column on every price list.
Compaction loops. When a thread nears the context ceiling, Codex summarizes and continues. That loop can misbehave: issue #35226 (July 24, 2026) reports auto-compaction rereading files, losing progress, and consuming paid credits on a Pro plan at 93% context. If you see frequent compaction warnings, start a brand new session immediately instead of continuing the thread.
Tool overhead. Issue #14750 (March 15, 2026) traces fast burn to oversized uncached prompt suffixes and repeated session polling on the UnifiedExec path. Your diff will never show this, because it is not in the diff.
Parallel agents multiply all four. Each subagent carries its own context, so a fan-out of five is five working sets rather than one.
Exit 1: spend the reset
This is the free exit, and the only one that does not change your bill. Run /status before the next long task and read two numbers: what is left on each limit, and when the next reset lands.
If your account carries earned resets, the CLI offers Redeem usage limit reset. They are one-shot credits against the clock, so spend them on a deadline rather than a refactor you could split.
For the rest of the window, downshift. Lower the reasoning effort, pick the smaller model, cut parallel agents to one, and start a new thread when a long one approaches the context ceiling. Compaction is a recovery path, not a free one.
Exit 2: buy headroom
Three paid levers exist: credits on workspace plans, a Pro upgrade, and paid resets of the weekly limit, which OpenAI documents in its help center (paid weekly rate limit resets).
Buying is rational when a blocked afternoon costs more than the reset: a demo, a deadline, a production fix. It is irrational as a habit, because the subscription was priced for average use, and the meter does not know you are in a hurry.
The trap is buying credits to keep a looping session alive. If a thread has crossed 80% of its context and is rereading files, more credits buy more of the same loop. Start a new thread first, then decide.
Exit 3: put the session on a metered bill
Log in with an API key and the billing model changes. With an API key configured, the CLI says Codex "will use usage-based billing with your API key." The five-hour and weekly plan caps stop applying. Your ceiling becomes the provider's rate limit and your balance.
There are two routes. An OpenAI API key keeps the same models at list price. A custom provider opens the field to other models. Since Codex 0.137.0 the custom-provider route requires an endpoint that speaks the Responses API natively; Chat Completions-only proxies fail at startup, which is why older tutorials break on current releases.
The working config for Kimi K3 on our endpoint:
model = "kimi-k3"
model_provider = "wallaby"
[model_providers.wallaby]
name = "Wallaby"
base_url = "https://api.wallabytoken.com/v1"
env_key = "WALLABY_API_KEY"
wire_api = "responses"
Save it as ~/.codex/wallaby.config.toml, export the key, and launch with codex --profile wallaby. The full walkthrough, including the metadata warning and the first-turn delay, is in our Codex CLI setup guide.
The cost shape differs in one way that matters: cached input. Kimi K3 costs $2.70 per million input tokens, $0.27 for cached input, and $13.50 per million output tokens on Wallaby, which is 10% off the official list. Agent loops resend context, so the cached column often decides the real total rather than the headline input price.
This exit is not automatically cheaper. If you exhaust the weekly cap every week, a subscription is probably the better deal. Metered billing wins on spiky usage, on teams that want one bill across tools, and on anyone who wants an open-weight model they can also run somewhere else.
Which exit fits
A blocked afternoon before a deadline: buy the reset or the credits, then fix the session that caused it.
Hitting the five-hour wall most days: fewer parallel agents, lower effort, new threads. If the wall still arrives, move to metered billing.
Hitting the weekly wall: do the math on last month before you upgrade. Metered tokens and a plan upgrade are both honest answers, and they suit different usage shapes.
Running a team: workspace credits with per-key limits on the provider side keep one runaway session from eating the month.
What you built
You now have a triage rule. Read the banner, run /status, and separate the two clocks before spending anything. The expensive mistakes are not the resets you wait for. They are the loops you fund.
Where to go next:
- The setup walkthrough: Kimi K3 in Codex CLI, including the warnings and the first-turn delay.
- The price comparison: Kimi K3 pricing across 18 providers, blended for a coding-agent workload.
- The migration angle: what replaced GPT-5.5 in Codex, if your default model changed in October.
If you want to test the metered route before committing, a new Wallaby Token account comes with $0.50 in trial credit, enough to run a few real Codex turns. The minimum top-up is $20, and the rate stays 10% off the official list.