Speed just went on the rate card
A mobile crane on hire costs the same whether it is lifting or standing. A monthly cost report does not care if it waits until Monday. Every job runs both kinds of work at once. Until this summer AI was sold at one price and one speed, so nothing on the invoice ever told the two apart. OpenAI has now split it into three things you choose between: where the work runs, how fast the answer comes back, and how much thinking you are paying for.
- 14x
- how much faster the top model now runs, for a few customers
- 2x
- what a faster answer adds to the bill
- 39x
- between the cheapest way to read a thousand documents and the dearest
The work stops living on one person's laptop
What happened. OpenAI is buying a company called Ona, announced on 11 June. Ona runs AI tools inside a firm's own secure cloud, so a job carries on after somebody shuts their laptop. OpenAI says the work worth doing now runs for hours or days rather than minutes. More than five million people use its Codex tool every week.
Why it matters. Take the lifting file on any decent sized job: the lift plans, the thorough examination certificates for every chain, sling and shackle, the RAMS, the appointed person's sign offs. Plenty of people have already pointed AI at that pile to sort it, check what is missing and chase it. It works. Then it turns out the whole thing runs on one person's laptop. It stops the week they are on leave. The record it built walks out of the door when they change firms.
The list OpenAI is selling this on is the list you would put in a subcontract: where it runs, what it is allowed to see, who holds the logins, what gets recorded, and who signs the output off. Those questions are not new. It is only new that you can now ask them of an AI tool and get a straight answer.
Do this. Write down every AI job your team leans on. Next to each one, put the machine it runs on and the name of somebody else who can start it again. One name in both columns is what to fix before you spend a penny on speed.
You can now pay extra for a faster answer
What happened. On 13 August OpenAI showed Ultrafast, its top model running up to 14 times faster than normal, which is quicker than anyone can read it. That one is with a handful of customers and not on sale. What you can buy today is Fast mode. Through the API it costs twice the normal rate. Inside Codex it burns credits two and a half times as fast.
| What you are buying | How fast | What it costs | When it is worth it |
|---|---|---|---|
| Batch | Slower, queued | Half the normal rate | Nobody is waiting |
| Standard | Normal | The rate card | Most work |
| Fast mode, API | Up to 2.5 times | Twice the rate | Someone is waiting on the phone |
| Fast mode, Codex | 1.5 times | Credits go 2.5 times faster | Someone is sat watching it |
| Ultrafast | Up to 14 times | Not published | Not on sale yet |
OpenAICodex speed settings, read 25 August
Why it matters. Here is where that choice gets made. The crane is on site, the slinger is booked, and the unit that turns up is not the one on the lift plan. The appointed person needs to know what the plan and the RAMS said about that load at that radius. Until they do, the crane, the gang and the delivery are all stood there. That is charged by the hour and it is the sort of hour a project manager remembers for a year.
That answer is worth paying double for. The same tool reading through the lifting file overnight to tell you which certificates expire next month is not. That one can queue, and it runs at half price while everybody is asleep.
Do this. Name the two jobs in your week where somebody is genuinely stood waiting on an answer. Pay for speed on those two. Put the rest on the overnight rate and stop paying a premium for work nobody is watching.
The cheap model got 80 per cent cheaper
What happened. On 30 July OpenAI cut the price of Luna, its cheapest model, by 80 per cent, and Terra, its middle one, by 20 per cent. The day before, it explained where the money came from. It used its own top model to rewrite the software that runs the models. That took 20 per cent off the cost of running them, and made them produce text more than 15 per cent more efficiently.
OpenAIAPI pricing, flagship models, read 25 August
Why it matters. Count what is in one lifting file, or in a subcontractor's O&M pack, and a thousand documents stops sounding like a lot. Reading that pile costs about $29 on the cheapest model and about $1,120 on the dearest one in a hurry. Same documents, same work, same answer at the end of it.
Which one you need depends on the job, not the budget. Pulling the expiry date off a certificate and filing it under the right crane is well defined work, and the cheap model does it. Reading a lift plan against a changed load, or a subcontract against its amendments, is not well defined, and that is what the expensive one is for. Most business cases were priced on one number, before July, and nobody has been back to them since.
Do this. Take one job you turned down on cost, price it again at today's rates, and check which model it actually needs. If the answer is only ever checked by a person anyway, it does not need the dear one.
What to ask before you sign
An AI quote now has three questions in it: where the work runs, how fast the answer comes back, and how much thinking you are paying for. Those are the same three things you would pin down on any package before you let it start, and they belong in the paperwork for the same reason.
If somebody is asking you to approve an AI tool this quarter and the paper in front of you does not answer all three, that is a twenty minute conversation.
Sources
Every figure and price in this edition was read from the page linked below on 25 August 2026. Where a number could not be confirmed at its publisher, it is not in the edition.
- OpenAIOpenAI to acquire Ona
- OpenAIPreviewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed
- OpenAIAdvancing the price-performance frontier with GPT-5.6
- OpenAIHow GPT-5.6 fuses frontier intelligence with frontier efficiency
- OpenAIAPI pricing
- OpenAICodex speed settings
- OpenAIChatGPT is now a partner for your most ambitious work
Followed this week
Where the week was picked up. Every figure and price above was then checked at the publisher and drawn from that source, so no chart here is traced off a video.
