How much Codex does ChatGPT Pro $100 really give you?

In our logs, a full Pro $100 weekly window priced at a median $175 of API-equivalent usage, between $100 and $234. That is about 2.6 times a $67 Plus week, not the advertised 5 times.

How fast can you use up a Codex weekly limit?

In 7.2 hours, in our heaviest window on Pro $100. Our six full windows lasted between 7 and 94 hours.

What is a banked Codex reset?

A one-off refill OpenAI grants that you redeem later. Redeeming it refreshes your Codex usage windows and moves your weekly reset date. Redeem it only when you are actually blocked, because a global reset can wipe the unused part.

How many Codex resets has OpenAI given?

Public trackers have logged dozens since 2025. In our 12-day test we used seven: two global resets, on October 2 and October 7, 2026, and five banked resets.

Is ChatGPT Pro $100 worth it for heavy Codex use?

Yes, at current reset rates. It delivered about $1,210 of API-priced Codex in 12 days, about 30 times its prorated cost. Without resets it would have been at most about $351, which is still several times the price.

We got 8 Codex weekly windows in 12 days on ChatGPT Pro $100: the reset log

Updated October 9, 2026 · first published October 9, 2026

Quick answer: Between September 27 and October 8, 2026, we ran Codex on ChatGPT Pro $100 as hard as it would go. In 12 days we opened 8 weekly usage windows and drained 6 of them. Two global resets and five banked...

Between September 27 and October 8, 2026, we ran Codex on ChatGPT Pro $100 as hard as it would go. In 12 days we opened 8 weekly usage windows and drained 6 of them. Two global resets and five banked resets kept us going. The shortest "weekly" window lasted 7 hours. All together the plan delivered about $1,210 of API-priced Codex, about 30 times the $40 that 12 days of a $100 plan cost.

The window itself was smaller than the label. A full Pro $100 week priced at a median $175. That is 2.6 times a Plus week, not the advertised 5 times. The resets, not the plan, are what made it a bargain.

Measured multiplier: 2.6x Plus, which is 53% of the 5x label.
Resets used: 7 in 12 days. Without them, 12 days would have held at most two windows, or about $351.
Biggest mistake: we redeemed a banked reset 12 hours before a global reset wiped it, which threw away about 86% of a window.

How we measured it

Codex writes a JSONL log of every session under ~/.codex/sessions/. Each request records cumulative token counts (input, cached input, output) and the current rate-limit state: plan type, the weekly used_percent, and the resets_at timestamp. We read about 1,300 session files and priced each request's token delta at the API list price of the model it used. We then grouped the requests by weekly window. When resets_at jumps forward before the old time has passed, a reset happened.

The logs show the plan changing from Plus to Pro $100 (prolite) on September 27. From then on only a weekly window exists, 10,080 minutes long. There is no 5-hour window, which matches OpenAI's pricing page.

The full log: 8 windows in 12 days

#	Window (UTC)	Lasted	Used	API value	Opened by
1	Sep 27 13:26 to 20:40	7.2 h	100%	$198	Upgrade from Plus
2	Sep 27 20:43 to Sep 28 14:45	18.0 h	100%	$159	Banked reset
3	Sep 28 15:06 to Sep 29 23:40	32.6 h	100%	$172	Banked reset
4	Sep 30 00:16 to Oct 2 08:48	56.5 h	100%	$100	Banked reset
5	Oct 2 08:49 to 21:10	12.4 h	14%	$32	Banked reset, then wiped
6	Oct 2 21:10 to Oct 6 18:59	93.8 h	98%	$175	Global reset
7	Oct 7 05:59 to 23:17	17.3 h	100%	$234	Global reset
8	Oct 8 06:30, still open	n/a	63%	$140	Banked reset

The six full windows priced at $100, $159, $172, $175, $198 and $234, a median of $175. The spread reflects the work. Window 4 was mostly GPT-6 Sol on long, cache-heavy sessions, while window 7 ran more GPT-6.1 Sol output.

Here is the same log as the weekly gauge, hour by hour. Every drop to zero is a reset.

Codex weekly gauge, hour by hour, Sep 27 to Oct 8 (B = banked reset, G = global reset)

The raw data: windows, hourly gauge and value, model mix. The math behind each number has its own write-up: the 2.6x multiplier, what the resets were worth, when to redeem a banked reset and the caching bill.

Pro $100 is not 5x Plus

Across 27 weekly windows since August, a full ChatGPT Plus week has priced at a median $67 (see how much Codex ChatGPT Pro includes). At 5x, a Pro $100 week should be worth about $335. Our measured median is $175, which is 2.6x. We are not the only ones to see this. On OpenAI's forum, a user who moved from Pro $100 to Pro $200 reported getting roughly half the effective usage they expected.

One caveat. About 6.4 billion of our tokens went to Luna, which is very cheap at API prices. If Luna draws on the weekly gauge at a higher rate than its API price implies, dollars understate how much quota we used. Even so, a window that a single heavy session can empty in 7 hours is not a week.

Where the tokens went

Model	Tokens	API value
GPT-6 Sol	3.4B	$565
GPT-6.1 Sol	3.2B	$510
Luna	6.4B	$95
GPT-5.5	33M	$32
Astra	10M	$16

Input totalled 13.1 billion tokens, and 97.6% of it was cached. Output was 47 million tokens. Agentic coding re-reads the same context over and over, so prompt caching is most of the economics. At uncached prices the same work would have cost many times more in API terms. GPT-5.5 retires on October 14, 2026, so its share drops to zero after that.

The reset that wiped a reset

A banked reset is a one-off refill OpenAI gives you to redeem later. According to OpenAI's help center, redeeming one refreshes your Codex usage windows and moves your weekly reset date. Global resets hit every account at once, with no warning. Public trackers have logged dozens of them since 2025 (see our Codex reset log).

On October 2 at 08:48 we hit 100% and redeemed a banked reset a minute later. Twelve hours later, with 14% of the new window used, OpenAI pushed a global reset at about 21:10 UTC. The global reset gave us a fresh window but did not credit the 86% we had not used, so the banked reset was mostly thrown away, about $150 of value. The global reset on October 7 at 03:35 was the second one inside five days.

How to get the most out of resets

Redeem banked resets only when you are blocked. One redeemed at 60% wastes 60% of a window.
Watch the reset trackers before redeeming. OpenAI has shipped global resets in clusters. If one landed recently or a launch is coming, use the banked reset late.
Put your heaviest work right after a reset. Long runs that need a lot of context burn the gauge fastest, so start them on a fresh window.
Keep your cache warm. At 97% cached input, breaking the cache with reshuffled context or a model switch partway through a session is the quickest way to burn a window.
Price your own window. Your logs already contain everything you need. Divide the dollars by the percentage points used and multiply by 100.

What this means after October 30

From October 30, Pro $200 drops from 20x to 10x Plus, the same per-unit price as Pro $100. If the $200 window scales the way ours did, it holds about $350 a week, not $670. That is less than Claude Max 20x delivers for the same money once you allow for Codex's per-task efficiency. The details are in our breakdown of the Pro $200 cut. Resets are generous, but OpenAI can stop them at any time. Treat them as a bonus, not part of your capacity plan.

Sources

Want this applied to your stack? Bring the provider bills, gateway logs, and top workflows; we will map the cost drivers and savings path. Book a free audit →

Back to finopsllm.com