We Used Codex Pro $100 for 30 Days: 14 Early Resets and $3,221 of API Value

Updated November 3, 2026 · first published November 3, 2026

In October we ran OpenAI's Codex on the Pro $100 plan and logged every request. The weekly limit refreshed early 14 times in 30 days, giving 15 weekly windows instead of the usual four or five. The usage was worth $3,221 at API list prices, about 32 times the $100 plan price.

Published 3 November 2026 · data 1 to 31 October 2026 (UTC) · method and caveats below

This is the month-long follow-up to our 12-day Codex log. Same method, same machine, now over 268,978 requests from 2,847 session files, 1 to 31 October 2026 (UTC). The figures are list-price equivalents, not what OpenAI charged, and we say below exactly what they leave out.

What surprised us

1. A reset about every two days. 15 windows in a month where the calendar gives you 4.3. The median window lasted 31 hours, not seven days.
2. Resets followed burn-outs. 9 of the 14 early resets came right after a window that had reached 95% or more. The logs cannot say whether each was an OpenAI-wide reset or a banked reset we redeemed; they only show the gauge jumping back before its scheduled time.
3. Almost everything was cache. 97.9% of the 37.0 billion input tokens were cached reads, which list at a fraction of the normal input price.
4. The cheap model did the volume, the expensive one did the damage. GPT-6 Astra made 1.1% of requests and 17% of the value. GPT-6 Luna made 59% of requests and 8.5% of the value.

The month, window by window

Window 1 is partial: our logs start on 1 October with the gauge already at 48%. Window 15 was scheduled to run to 4 November; its dashed line is cut at the edge of the chart.

What was a window worth?

A fully used window was worth $258 on average, but the spread is wide ($182 to $423). The biggest, window 15, ran for 70 hours and 39,755 requests. Value per percent of gauge is not constant, because it depends on the model mix and how much of the context was cached.

Why this is higher than our earlier $175 median

Our earlier multiplier article measured $100 to $234 per full window, median $175, over six windows between 27 September and 8 October. This month's 9 fully used windows ran $182 to $423, with a median of $237 and a mean of $258. It is the same method on different weeks and a different model mix, so the two sets are not directly comparable. We have not isolated one cause for the difference; the point is that a full window's list-price value is a range, not a fixed number.

What would a reset-free month have been?

Estimate, not measurement: with one natural window a week, a 31-day month holds 31 / 7 = 4.4 windows. At the average fully used window ($258), that is about $1,141. We measured $3,221, so the early resets made the month worth roughly 2.8 times a no-reset month, or about $2,080 more at list prices. Treat the $1,141 as a rough baseline: our window values varied by more than 2x.

Which models used the plan?

GPT-6.1 Sol carried 64% of the value. The lesson is the same one we see in API bills: request counts say little about spend. Count value by model, not requests. See cost per successful task for the measure that matters once you add retries.

How we measured it

• Source: Codex's local session logs. Each request records cumulative token counts and the rate-limit state (plan type, weekly percent used, scheduled reset time). The plan type read prolite throughout, which is Pro $100 (see the earlier log).
• Value: each request's token change priced at OpenAI's published short-context rates (inputs up to 272K tokens) from the OpenAI pricing page, checked 3 November 2026: GPT-6.1 Sol $2.00 input, $0.10 cached, $10.00 output per million tokens; GPT-6 Astra $10 / $1 / $50; GPT-6 Luna $0.10 / $0.01 / $0.50; GPT-6.1 Terra $1 / $0.05 / $5; GPT-6 Sol $2 / $0.20 / $10.
• Windows: a new window starts when the scheduled reset time changes. Mixed sessions sometimes report stale reset times, which produced 35 raw fragments; we merged fragments that shared a reset time into the 15 real windows above. An early reset is a window that opened more than five minutes before the previous one was scheduled to end.

What the numbers leave out

• 425 auto-review requests (0.16% of the total) use a model with no published price and are counted as $0.
• Requests over 272K input tokens would price higher than the short-context rates used here.
• 326 log lines could not be parsed, and the logs end on 31 October at 05:56 UTC.
• List-price value is a way to size usage. It is not an invoice, a saving, or a promise that later months will match; OpenAI sets plan limits and can change them.

What to take from it

1. Measure per window, not per month. A monthly average hid windows from $182 to $423.
2. Log the gauge and the reset time with every request. It is the only way to see an early reset after the fact.
3. Watch cache share and model mix. They moved the value more than request counts did. For the mechanics, see Codex prompt caching cost.
4. Plan the reset, do not count on it. Read when to redeem a banked reset and the dated reset history before you build a workflow around one.

Related

• 8 Codex weekly windows in 12 days on Pro $100
• Codex reset history: every OpenAI reset, dated
• Codex banked reset: when to redeem it
• ChatGPT Pro $100 promises 5x Plus for Codex. We measured 2.6x
• 98% of our Codex input tokens were cached

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