Codex Pro 200 Quota Test: GPT-6 Astra and GPT-6.1 Sol Weekly Tokens

Timeline (Singapore time, UTC+8)

2026-09-29 14:41 / 23:10: Tibo of OpenAI (@thsottiaux, X bio "Codex & ChatGPT @OpenAI") posts on X that Pro 200 will reopen with a changed usage calculation (posts in Sources).

2026-10-28 14:25–14:35: aiplans first retest (28-call Astra main batch); report merged into the site repo at 14:43.

2026-10-28: Pro 200 price of US$200/month verified on aiplans.

2026-10-29: gpt-6-astra API pricing snapshot verified (the only price basis for USD figures here).

2026-10-30 00:56: Tibo announces on X the Pro 100/200/500 and Business Premium allowance change, "effective today" (posted 2026-10-29 16:56 UTC).

2026-10-30: checked the OpenAI Help Center articles "About ChatGPT Pro tiers" and "Managing usage with GPT-6 Astra in Work and Codex".

2026-10-30 14:36–14:51: aiplans current retest (60-call main batch each for Astra and Sol); report merged at 14:59.

2026-10-30: review date for external sources on this page.

Key results: weekly capacity for Astra and Sol

GPT-6 Astra: all 60 main-batch calls succeeded, 5,580,339 known tokens, weekly usage 49% → 53%. Point estimate ~139.5M tokens/week (139,508,475). Treating each whole-number reading as ±0.5 percentage points gives a range of ~111.6M–186.0M (111,606,780–186,011,300). This is a rounding range, not a statistical confidence interval.

GPT-6.1 Sol: 59 of 60 main-batch calls succeeded (one "Selected model is at capacity" error), 5,484,515 known tokens, weekly usage 48% → 49%. Only a conditional lower bound is available: ~274.2M tokens/week (274,225,750). There is no upper bound and no point estimate.

USD equivalent: at the standard gpt-6-astra API list price (USD per 1M tokens: input $10 / cache read $1 / output $50; pricing snapshot verified 2026-10-29, not re-verified for this test), Astra's test workload is worth about $1,222 per week, range $978–$1,630 (rounding only). This is an API list-price equivalent, not a plan balance, and it does not mean OpenAI deducts quota at API prices. No USD equivalent is given for Sol.

Pro 200 costs US$200/month (verified on aiplans 2026-10-28); see OpenAI pricing & subscription plans.

What these numbers mean, and when they apply

ChatGPT Pro 200 only. The Pro 200 plan was confirmed by the owner, not by independent billing verification; the upgrade is owner-reported. The account’s allowance tier after September 29 has not been verified, nor has the effect of the October change on this account. The measurements neither prove nor disprove the official statement.

Astra and Sol share one weekly pool. If you use both in the same week, you draw down one pool; the total is not 139.5M + 274.2M.

The Codex CLI reports only the requested model name; it cannot prove which model actually served the request.

Only the weekly (10,080-minute) window was readable. The 5-hour and monthly windows were not returned, and this page does not extrapolate them.

These are conditional estimates for one workload mix: ~93K input tokens per call, short output, high concurrency. Cache reads were about 14.45% of Astra's main batch and output was minimal. A different input, cache and output mix will yield a different token count.

The estimates assume linear, settled metering and no other traffic. They are not an OpenAI commitment.

Why Sol is only a lower bound

During Sol's main batch the weekly meter moved only one point (48% → 49%), the controller session counter also increased, and one request failed with an at-capacity error that returned no usage. The report does not attribute the whole change to Sol, nor does it use API prices to subtract other usage. Instead it takes the most conservative view: with whole-number readings, the true increase was at most 2 points, so Sol's known usage consumed no more than 2 points. Weekly capacity is therefore at least 5,484,515 ÷ 2% ≈ 274.2M tokens.

The bound relies on four assumptions: settled metering, linear consumption, no refunds, and no reset or quota change. If any fails, the figure cannot be used. It shows a floor, not Sol's full weekly quota.

How this compares with the October 28 run

In the first retest on Oct 28 (run 14:25–14:35, before OpenAI's announcement), Astra's 28-call main batch gave a point estimate of ~86.5M tokens/week, range ~64.9M–129.8M; Sol's readings were inconsistent and no capacity was published.

This run's Astra point estimate is higher, but the two ranges overlap (the new lower edge, 111.6M, is below the old upper edge, 129.8M), and the cache and system-input mix differed between runs. So this result does not prove how much the quota changed, and the ratio of the two point estimates cannot be used as an upgrade multiplier.

Official statements: OpenAI Help Center

Official statements: OpenAI Help Center (About ChatGPT Pro tiers, checked 2026-10-30):

"Pro 100 and Pro 500 now include twice their previous Work and Codex allowance. On Pro 200, only the lower allowance doubled. New subscriptions include the updated allowance, and existing subscribers receive it automatically without waiting for renewal."

"The reduction announced in our September 29 email is canceled. If you already had the higher Work and Codex allowance, you keep it. If you had the lower allowance, it has doubled to match. Your subscription price stays at $200/month."

"A larger allowance does not guarantee twice as many completed tasks."

Official statements: Tibo on X

Official statements: Tibo on X (original post, 2026-10-30 00:56 SGT):

"We are doubling all usage limits for all Pro 100/200/500 tiers and Business Premium, effective today."

"We heard your kind feedback and won’t be reducing usage limits for Pro 200, so your Work and Codex usage will stay the same."

Tibo's post is less specific about Pro 200 than the Help Center; for Pro 200 we follow the Help Center rule that only the lower allowance doubled.

Official statements vs measurements: what matches

Timing: one of our runs was before the announcement (Oct 28, 14:25) and one after (Oct 30, 14:36).

Weekly window only: the Help Center says "Pro 100, Pro 200 and Pro 500 currently have no five-hour usage limit in Work and Codex.", consistent with us reading only a weekly window.

Shared pool: the Help Center says "Work and Codex share the usage allowance included with your plan." and "Switching models does not restore allowance in a shared usage pool.", consistent with Astra and Sol drawing on one weekly pool in our test.

Direction: Astra's point estimate is higher than on Oct 28, which does not contradict "no reduction".

Official statements vs measurements: what cannot be shown

Multiplier: the official doubling applies only to Pro 200 accounts that had the lower allowance. The account’s original allowance tier after September 29 and before this change has not been verified, so we can't even say what OpenAI's rule implies for it.

Precision: the old and new Astra ranges overlap, so "unchanged" can't be ruled out; with whole-percent readings, doubling can't be ruled out either. Our data neither proves nor disproves the official doubling, and this page gives no multiplier of its own.

Sol: no Sol capacity was published on Oct 28, so there is no baseline to compare.

OpenAI does not publish Pro 200 token allowances; the figures here are measured estimates for one workload, not official quotas.

Method

Tool: Codex CLI signed in with the ChatGPT Pro 200 account, reading that account's weekly quota window.

Workload: the same synthetic Python code-review prompt as Oct 28, ~93K input tokens per call, short output; a random nonce per call, with caching taken from actual usage. No tool execution was observed.

Batches: per model, 4 calibration calls, a 32-call concurrent extension, then a 60-call main batch (up to 12 concurrent). Capacity uses the main batch only; batches are not merged across models.

Readings: weekly usage before and after each batch, with immediate and delayed end readings at least 10 seconds apart, plus checks on reset time and controller counters. Collection time is not server settlement time.

Counting: ordinary input = input_tokens − cached_input_tokens; total = input_tokens + output_tokens, without double-counting reasoning; cache writes were 0.

Calculation: Astra point estimate = total tokens ÷ 4%, range = total ÷ 5% to total ÷ 3%; Sol lower bound = known Sol tokens ÷ 2%.

Stop rule: testing stopped at about 53% used; calls were not added indefinitely after errors.

The per-call ledger and all window readings are in codex-models-2026-10-30.json.

Subscription or API: how to use these numbers

First separate the two billing paths: Codex signed in with a ChatGPT plan uses the plan's allowance; Codex signed in with an API key is billed at API prices, and a subscription does not offset API charges (see OpenAI's Codex pricing).

1. Measure the tokens you actually use in Codex in a week, split into ordinary input, cache reads and output.

2. Price that usage at API rates: ordinary input ÷ 1M × input rate + cache reads ÷ 1M × cached rate + output ÷ 1M × output rate. For Astra you can use the 2026-10-29 snapshot above; for other models, check the current price on the model page.

3. Compare with the plan: Pro 200 is US$200/month. Price a month of your real usage as in step 2 and compare it with the monthly fee. Well below the fee, pay-as-you-go API may cost less; well above it, the subscription is better value, but check whether your weekly usage would exceed the weekly capacity on this page. For reference, at the 2026-10-29 pricing snapshot, a full week of this Astra test workload is worth about $1,222 at API list price ($978–$1,630), not a plan balance or deduction rule.

4. Check whether your mix is similar: this test used long input, short output and low caching. With more output or more caching, both the usable token count and its API value change, so do not apply these figures directly.

5. If you mix Astra and Sol, budget against one pool; do not add the two capacities.

6. Hitting the limit does not force an upgrade: per OpenAI, Plus and Pro users who reach their limit can buy additional credits, or run extra local chats with an API key at standard API rates. Check credit rates on the official page; this guide does not estimate them.

Your own account's usage meter and OpenAI's official documentation take precedence. Compare channel API prices on the GPT-6 Astra and GPT-6.1 Sol model pages, and plans in the subscription comparison.