Ultrafast time and usage calculator
Work out how much sooner a Codex or ChatGPT Work job would finish on Ultrafast or Fast mode, from its time on Standard and how much of that is writing, and how much more of your allowance it would use.
- Time on Ultrafast
- 3 min 27 sec
- You’d save about 6.5 minutes and use 8 times the allowance.
- Time saved
- 6 min 33 sec
- from 10 min on Standard
- Whole job speed-up
- 2.9x
- Even instant writing couldn’t beat 3.8x
- Allowance used
- 8x Standard
- With bought credits: 6x
The new time is the Standard time × ((1 − the writing share) + the writing share ÷ the speed-up), which is Amdahl’s law. It counts everything except writing, like thinking, waiting, running commands and checking the work, as taking the same time in both modes, and it assumes the job writes about the same number of tokens either way. In both our agent runs Ultrafast also spent less time thinking and waiting, so the real job finished sooner than this predicts.
The writing speed-ups are our own timings of single replies from GPT-6 Astra in Codex on 30 September 2026; yours will vary. OpenAI’s Codex pricing, checked the same day, has Ultrafast using included usage at 8x the Standard rate and bought credits at 6x, and Fast at 2.5x and 2x.
What goes in
Enter how long a job takes on Standard, and roughly how much of that time the model spends writing. In our code-heavy agent run, Standard spent 74% of its time writing (294.1 of 396.6 seconds), and the rest went on thinking and waiting, and on running commands and checks. A quick reply with no commands is close to 100% writing.
Then pick the mode and the kind of writing. The Ultrafast speed-ups are our single-reply timings on GPT-6 Astra: 4.5x for a story, 8.7x for code and 14.8x for the numbers one to four hundred in words. OpenAI’s own figure is “up to 8x”. For Fast, Codex’s model list says 2x. You can type your own figure instead.
What the result means
Only the writing gets faster, so the new time is the Standard time × ((1 − the writing share) + the writing share ÷ the speed-up). That’s Amdahl’s law. A 10-minute job that’s 74% writing, with code written 8.7 times faster, takes about 3 minutes 27 seconds on Ultrafast. That’s 2.9 times faster and saves about 6.5 minutes. Even if the writing took no time at all, the job couldn’t get more than 3.8 times faster.
Allowance used is how fast the job uses up your plan compared with Standard, assuming it writes about the same number of tokens either way. Ultrafast uses included usage at 8 times the Standard rate and bought credits at 6 times. Fast uses 2.5 times and 2 times.
The sum treats thinking, waiting and commands as taking the same time in both modes. In our runs Ultrafast spent less time thinking and waiting too, so the real jobs finished sooner than it predicts: our first agent run came out 3.5 times faster, where the formula says about 3.
The figures behind it
Ultrafast is in Codex and ChatGPT Work on the $500 Pro plan (not the $100 or $200 ones) and eligible Enterprise and Edu plans, and in the API at API prices. The usage rates come from OpenAI’s Codex pricing page, checked on 30 September 2026.
Our single-reply speeds are the middle result of three rounds, after taking off the usual wait before a reply starts, estimated from a one-word reply in the same round. Standard wrote at 31 to 33 tokens a second whatever the text. Ultrafast wrote at 140 on a story, 283 on code and 491 on counting. These are our timings on one day, and yours will vary.
Sources
- OpenAI Codex docs: Speed and Pricing
- Codex’s model list for GPT-6 Astra’s Fast mode: models.json in the openai/codex repository
- Hill and Marty, July 2008: Amdahl’s Law in the Multicore Era, IEEE Computer
- Gene Amdahl, 1967: Validity of the single processor approach to achieving large scale computing capabilities