Measured on October 19, 2026 · GPT-6 Astra · $200 Pro plan · API and credit prices checked November 3, 2026
A weekly allowance on the $200 Pro plan is worth roughly 425 million Codex tokens. At API prices, that is about $1,271 of work, for similar coding tasks.
OpenAI doesn’t publish these numbers, so I worked them out from my own usage. I was running real coding tasks in 20 Codex sessions at the same time, and the weekly meter climbed from 3% to 90% in 132 minutes. The whole batch finished in under three hours. That is about a week of allowance, used up in an afternoon. Limits and prices change, so the date above matters.
What I ran
Twenty coding sessions in parallel, on the gpt-6-astra model with reasoning effort set to xhigh. The main batch took 2h 58m 34s. Together with a small pilot run, the counters recorded 452,810,013 tokens in 469 sessions. Summed session time was 54h 03m, but the two batches took only 3h 17m of real time.
How I got to 425 million
Between the 3% and 90% readings, finished sessions used 357.51M tokens and the conversation added 0.30M. Sessions still running added an estimated 10.41M, split by elapsed time. That is 87 percentage points of allowance, so:
(357.51M + 0.30M + 10.41M) ÷ 0.87 ≈ 423.24M tokens
The earlier 3% to 45% stretch gives 431.30M, which agrees. I round to 425M and call 400–450M a practical range.
Why the token count is so big
Cached tokens were 88.4% of all input. Codex keeps re-reading the same context, and every re-read counts. Output is only 2% of the tokens, but as the chart shows, it is a third of the cost.
What it is worth
At base Standard API rates, 425M tokens of this mix come to about $1,271. The rates, per million tokens, are $10 uncached input, $1 cached input and $50 output (OpenAI’s Astra pricing).
One catch: a request with more than 272,000 input tokens is billed at $20 / $2 / $75, for the entire request. I only kept session totals, not per-request sizes, so here is the 452.81M-token run under three assumptions about how much of it was that large:
| Share of each token type in >272K requests | API-equivalent cost |
|---|---|
| 0% | $1,353.93 |
| 25% | $1,636.69 |
| 100% | $2,484.96 |
In credits, OpenAI’s published rates are 250 credits per million uncached input tokens, 25 for cached input and 1,250 for output. With 2,500 credits = $100, a credit is worth about 4¢. For my mix that works out to 74.75 credits, or $2.99, per million tokens. One credit buys about 13,378 tokens.
The fine print
This is one $200 Pro account and one kind of work, so treat it as a ballpark, not a guarantee. It is also a snapshot of October 19, 2026, and OpenAI can change allowances and prices. I did not keep the exact moment the meter hit 100%, so “about a week in three hours” is an estimate from the meter readings, and the 400–450M range is not a statistical confidence interval. The dollar figures are what the tokens would cost at API rates, not a bill. My credit balance did fall by 1,308.25 credits (about $52.33) after the run, but that is account-wide and includes subscription-funded work, so it can’t be turned into an exchange rate.
What I would budget
For a similar run, plan on around 400M tokens as a cautious start and watch the real meter. I was not trying to burn tokens, I was doing actual work, so next time I will also keep per-request input sizes and more allowance snapshots. With 20 sessions running at once, those are painful to reconstruct afterwards.
P.S. Others measured the same thing
I only went looking for other estimates after I had written this article, so they did not influence it. Three careful independent measurements land close to my number:
- Yad Konrad went through 767 session logs. His September usage on Astra came to about $1,260 per full week of the $200 plan.
- Inventive HQ measured about $13 per 1% of the meter on the old Pro 200 allowance, which is about $1,300 per week. Mine was $12.71 per 1%.
- SemiAnalysis (Hacker News discussion) put the reduced Pro 200 allowance at about $669 per week. Doubled back to the old level, that is about $1,340.
Two things stand out from reading them.
The meter seems to count dollars, not tokens. These measurements find that the weekly meter tracks API list-price dollars, so the dollar figure is the number that carries over between people. That would also explain why some report billions of tokens per week while I measured 425 million: they mostly used the cheaper Sol model and had 96–97% of their input cached, against my 88%.
The numbers are a moving target. OpenAI’s Pro help page says the reduction announced on September 29 is canceled. Anyone who had the higher allowance keeps it, and those on the lower one get it doubled to match. Some of the measurements above were taken during the reduced period. Also, from December, ChatGPT and Codex will share one allowance measured in tokens, so this will need measuring again.