Codex Usage Limits (November 2026): ChatGPT Plus & Pro Weekly Limits, Measured
Measured over 2026 11/09 – 11/16, a ChatGPT Plus week of Codex is worth about $47 of official API usage, 10× its price, about 13.4M tokens excluding cache reads. OpenAI changed its limits on 2026-09-29; only data after that change is used. ChatGPT Pro $100 (Codex) (1 accounts) is still collecting; published at 5 accounts.
Data updated 2026-11-16
How much Codex do ChatGPT Plus and Pro give per week?
Codex on ChatGPT plans also has a 5-hour limit and a weekly limit. A full week, measured on real accounts (data after the 2026-09-29 change):
- ChatGPT Plus (Codex) ($20/mo): ≈ $47 of API use a week (≈ $202/month) · 13.4M uncached tokens · ≈ 6.3 full 5-hour windows
- ChatGPT Pro $200 (Codex) ($200/mo): ≈ $736 of API use a week (≈ $3,191/month) · 141M uncached tokens · 16× ChatGPT Plus (Codex)
Did OpenAI reduce Codex limits?
No cut measured. In the week of 2026 11/09 – 11/16, a full week on ChatGPT Plus (Codex), ChatGPT Pro $200 (Codex) held about the same as in the week of 2026 11/02 – 11/09. The last official change was on 2026-09-29: OpenAI halved the ChatGPT Pro $200 allowance (from 20× Plus to 10×) and added a $500 tier; existing subscribers keep the old limits until Oct 29.
- ChatGPT Plus (Codex): No change · uncached tokens 9.4M → 13.4M/wk (+42%) · API value $46 → $47 (+1%) · from the model mix and output share
- ChatGPT Pro $200 (Codex): No change · uncached tokens 90.4M → 141M/wk (+56%) · API value $581 → $736 (+27%) · within sampling error
Measured weekly limit by plan
| Plan | Price | Weekly value | × price | Non-cached tokens / wk | Output / wk | 5-hour window | Accounts | Week |
|---|---|---|---|---|---|---|---|---|
| ChatGPT Plus (Codex) | $20/mo | $47 ($0.13–$58), ≈ $202/mo | 10× | 13.4M | 1.3M | $7.5 (≈ 6.3 a week) | 5 | 2026 11/09 – 11/16 |
| ChatGPT Pro $200 (Codex) | $200/mo | $736 ($508–$1,012), ≈ $3,191/mo | 16× | 141M | 9.9M | — | 6 | 2026 11/09 – 11/16 |
| ChatGPT Pro $100 (Codex) | $100/mo | Collecting (1 accounts; published at 5) | ||||||
Accounts on the same plan differ too: the middle half of ChatGPT Plus (Codex) accounts sits between $0.13–$58 (±100%); the middle half of ChatGPT Pro $200 (Codex) accounts sits between $508–$1,012 (±37%). Vendors A/B-test limits on a few accounts (SemiAnalysis found one of three identical accounts about 20% lower); we measure dozens to hundreds of accounts per plan, so a single outlier does not move the median.
Compared with SemiAnalysis
SemiAnalysis measures with controlled tests what one model on a fixed workload can be worth at most; we measure what a plan is worth in real use. The two agree on scale; the FAQ below explains the gap.
| SemiAnalysis: ChatGPT Plus, GPT-6 Astra only | $162/mo |
|---|---|
| SemiAnalysis: ChatGPT Pro $200, GPT-6 Astra only | $2,897/mo |
| SemiAnalysis: ChatGPT Pro $500, GPT-6 Astra only | $6,955/mo |
| Wokey measured: ChatGPT Plus, real traffic (5 accounts) | $202/mo |
Why "how many tokens per week" has no single answer
Most tokens in a long session are cache reads, and cache reads barely count against a plan. The same weekly limit can carry several times more total tokens for someone with a high cache hit rate. Compare plans by tokens excluding cache reads, or by API value, never by total tokens.
- ChatGPT Plus (Codex): 89% cache reads, 136M tokens a week, of which only 13.4M are not cache reads
- ChatGPT Pro $200 (Codex): 92% cache reads, 2.08B tokens a week, of which only 141M are not cache reads
Subscription or pay-as-you-go
A plan is the best deal when you use all of it. Most people do not: below the share shown here, paying per token costs less than the subscription. When you hit the weekly limit, you can keep using the same model on Wokey.
- ChatGPT Plus (Codex): if you use less than 98% of its weekly limit, Wokey pay-as-you-go costs less
- ChatGPT Pro $200 (Codex): if you use less than 67% of its weekly limit, Wokey pay-as-you-go costs less
Week over week
Remeasured every Monday. When a vendor changes its limits, these lines move; changes under about 10% in a single week are mostly sampling noise.
- ChatGPT Plus (Codex): 10-08 $92 → 10-19 $58 → 10-26 $55 → 11-09 $46 → 11-16 $47 (No change)
- ChatGPT Pro $200 (Codex): 11-02 $581 → 11-16 $736 (No change)
Limit changes
- 2026-09-29: OpenAI halved the ChatGPT Pro $200 allowance (from 20× Plus to 10×) and added a $500 tier; existing subscribers keep the old limits until Oct 29 (SemiAnalysis)
"You've hit your usage limit"
Codex prints this when the plan's 5-hour or weekly limit is used up. The rest of the line says what you can do:
Try again at …: Wait until the time shown; the limit resets on its own.Visit https://chatgpt.com/codex/settings/usage to purchase more credits: Your plan can buy credits to keep going before the reset.Upgrade to Plus to continue using Codex: You are on the free plan.To get more access now, send a request to your admin: A Business, Enterprise or Edu workspace: credits are managed by your admin.You've hit your usage limit for <model>. Switch to another model now: Only that model's limit is used up; other models still work.
Other clients show the raw response instead: {"error":{"type":"usage_limit_reached","message":"The usage limit has been reached","plan_type":"plus","resets_at":…,"limit_window_minutes":300,"resets_in_seconds":…}}. resets_in_seconds is the time left until the reset (12499 is about 3.5 hours), resets_at is the same moment as a Unix timestamp, and limit_window_minutes 300 means the 5-hour window ran out.
{"detail":"Rate limit exceeded"}
A 429 for sending too many requests in a short time, often from several Codex sessions or subagents running at once. It is not the plan limit: there is no reset time and your remaining usage is untouched. Codex retries on its own and, when every retry fails, prints "exceeded retry limit, last status: 429 Too Many Requests". Wait a minute and run fewer sessions in parallel.
How we measure
- The data comes from limit changes on real subscription accounts behind the Wokey gateway, together with the traffic they served. When an account’s weekly usage rises by at least 10 percentage points, the requests in that span, divided by the rise, give a full week of traffic.
- Each span needs at least 10 requests, with no more than 5% failed or boundary-crossing requests. A span is left out when its gateway traffic accounts for less than 70% of the limit it used, compared with the plan’s highest-reading accounts (the 90th percentile): that account was mostly used outside the gateway, and use outside the gateway only ever lowers a reading. Spans are first reduced to a median per account, then to a median across accounts, so every account counts once.
- Traffic is priced at each model’s official API rates, including cache reads and 5-minute and 1-hour cache writes. Cache reads barely count against the limit while uncached input and output weigh heavily, so the same limit converts to more dollars in a cache-heavy week. To keep weeks comparable, for Claude we fit each week how much of the limit each token type uses, then value every plan at one fixed traffic split: Claude Code traffic on Wokey in the week to 2026-11-02, about 92% of tokens cache reads. The dollar figure means "what a full week of that traffic would cost on the official API"; it is not the unit the vendor meters in. Claude Fable 5.2 uses noticeably more of the limit per official-price dollar than other models, so its weight is fitted separately and its traffic is converted to the dollars the same share of the limit buys at the fixed traffic split.
- Every Monday after 04:00 UTC the past 7 days are remeasured. A plan is published once at least 5 accounts are measured. If a value moves more than 15% from the previous week, it is checked by hand before the page updates.
- The limit is worth more or less depending on the model mix: the same limit spent on a different main model converts to a different API value. Each plan lists its model mix.
These are measured estimates, not figures published by the vendors. Vendors can change limits at any time, and use outside the gateway or delayed metering can shift the results.
FAQ
What is the ChatGPT Plus weekly Codex limit?
Measured, a ChatGPT Plus week of Codex is worth about $47 of official API usage (5 accounts, 2026 11/09 – 11/16), 10× its price and about 13.4M tokens excluding cache reads. OpenAI changed its plan limits on 2026-09-29; only data after that change is used.
Did OpenAI reduce Codex limits?
No cut measured. In the week of 2026 11/09 – 11/16, a full week on ChatGPT Plus (Codex), ChatGPT Pro $200 (Codex) held about the same as in the week of 2026 11/02 – 11/09. The last official change was on 2026-09-29: OpenAI halved the ChatGPT Pro $200 allowance (from 20× Plus to 10×) and added a $500 tier; existing subscribers keep the old limits until Oct 29.
How much Codex does ChatGPT Pro give?
ChatGPT Pro $200 (Codex): about $736 of official API usage a week (6 accounts). ChatGPT Pro $100 (Codex) (1 accounts) has too few measured accounts so far; published here at 5. OpenAI changed the Pro plan limits on 2026-09-29.
Why does my Codex limit drain so fast?
Codex meters by tokens, not messages: larger context, longer outputs and heavier models (such as GPT-6 Astra) all drain it faster. Plus is capped by both a 5-hour window and a weekly limit.
How do the weekly limit and the 5-hour limit interact?
Both apply at once: the 5-hour window caps bursts and the weekly limit caps the 7-day total, and whichever runs out first stops you. Higher tiers have large 5-hour windows, so the weekly limit is usually the one you hit.
Why do these Codex numbers differ from SemiAnalysis?
SemiAnalysis (2026-10-05) measured an agentic workload entirely on GPT-6 Astra: ChatGPT Plus about $162 a month, Pro $200 about $2,897 and Pro $500 about $6,955; since Sep 29 the three Pro tiers give the same value per dollar. We measure ChatGPT Plus at about $47 a week (≈ $202 a month), higher than their Astra figure because most of the value in real traffic comes from Sol-class models, and SemiAnalysis finds Plus gives more value per dollar on Sol than on Astra.
Where does the data come from, and how often is it updated?
From limit changes on real subscription accounts behind the Wokey gateway and the traffic they served, remeasured every Monday. Plans with too few accounts are marked as collecting and show no numbers.