Cheap Brains, Dear Minutes: How OpenAI Is Repricing Intelligence

OpenAI made its intelligence five times cheaper on Tuesday and its impatience six times dearer, and the two moves are the same idea.

Begin with the cheap half. GPT-6.1 Sol, launched at DevDay in San Francisco, costs $2 per million tokens going in and $10 coming out. (A token is a small chunk of a word, so a million of them is a long shelf of books.) The flagship, GPT-6 Astra, costs $10 and $50 on OpenAI’s pricing page. Sol is therefore a fifth of the price, and OpenAI’s own claim is that it delivers close to Astra’s intelligence for coding, for driving a computer and for ordinary professional work. On one computer-use test the company reports 71.4 per cent for Sol against 73.5 for Astra, at about a seventh of the cost per task.

I should say plainly that every benchmark figure here is OpenAI’s, and The Decoder notes that the company describes them as preliminary. Nobody outside has had long with the thing. But the price is not a claim. It is on the page.

Now the dear half. Switch on a new tier called Ultrafast and Astra itself costs $60 in and $300 out, six times standard. That is the sort of arithmetic a company does when it has decided where the value has gone. The cheap brain and the expensive minute are the two ends of one pricing strategy: intelligence is being marked down by the fortnight, and what OpenAI would now like to sell you is the time you save, the room you are given, and the place you stand in the queue.

The price list as a confession

Consider what the new price list admits. The Decoder observes that $2 and $10 is exactly what Anthropic charges for Claude Sonnet 5.5, and exactly half of what it charges for Claude Opus 5.5, which is $4 and $20. Sam Altman told CNBC that OpenAI has “the best models to offer at every price point.” It is a nice line, and also the sort you say when the price points are the thing being watched.

The detail that may matter more is the cache. When a program sends the same long preamble to a model again and again, the provider can remember it and charge less for each repeat. Sol’s cached input is $0.10 per million tokens, half of what GPT-6 Sol charged. Several commenters on the Hacker News thread, which had 573 points and about 490 comments a few hours after the keynote, called that halved cache price “the actual big announcement”. They may be right. It never makes a keynote slide, and it quietly rearranges what a busy program costs to run.

Here, then, is the awkwardness. Astra is the model OpenAI has just built its always-on dots agents upon, and the model whose next version, 6.1 Astra, it shelved on Monday. Sol arrived a week after the previous Sol, as TechCrunch and CNBC both stressed, and it is cheap. If Sol is even most of the way to Astra, most of the world has a good reason to stop paying for Astra. A product that eats its bigger sibling is a strange thing to launch on a day devoted to selling more.

The crowd, it must be said, was not swept away. On Hacker News many users said that GPT-6 Sol and Astra had underperformed Anthropic’s Opus 5.5 in real coding work, and were sceptical of benchmarks. Several noted, sensibly, that it is too early to judge 6.1. A recurring view was that DeepSeek V4.1 Flash and other open models, which anyone can download and run, are “good enough” for much of the work. These are anecdotes from strongly opinionated people, not a survey, but it is telling that the argument in the room was about value rather than wonder. Simon Willison, who was at the event, offered the driest verdict: his pelican-drawing benchmarks for 6.1 Sol were “not notably different from the GPT-6 family.” Some commenters speculated that Sol was a relabelled internal model, or a hurried answer to Opus 5.5. OpenAI has offered no evidence for either, and I would leave them where they are, as gossip.

Not everything is settled even on OpenAI’s side. Its page says Sol fails to tell you that a search tool was broken in 2.1 per cent of cases; The Decoder reports 2.8 per cent. I do not know which is right, and the gap is small, but a reader deciding whether to trust a cheaper model with unsupervised work might like the numbers to agree. Cognition, which put Sol into its Devin coding agent, reported on its own test that Sol scored 60.4 per cent against 60.7 for the earlier Sol, at $0.31 a task at medium effort. That is a lead, not a verified result, and it points somewhere interesting: much of the improvement is in the bill. Sol is also not yet in regular ChatGPT chat, only in the API, Work and Codex, which leaves the busiest door in the house closed for now.

Altman, asked about open models, said he does not think one beats OpenAI on a cost-price comparison. Sol at $2 and $10 is a fair attempt to make him right. Yet I do not think the move a mistake. The alternative is to leave the cheap end to others while defending a premium that customers are starting to doubt. Better to be your own competition than to let someone else have the job.

Six times the price for six times the speed

The other model on the bill is really a speed setting. Ultrafast, one of the less flashy announcements of the day, is OpenAI’s premium latency tier. The company says it gives up to 8x faster generation in Codex, its coding assistant, at up to 300 tokens per second, and up to 6x faster through the API, the plumbing developers use to build their own things. The API price is 6x standard.

Read that twice, because the shape is the story. It is not a premium product in the way a better ingredient is premium. It is a queue-jumping fee.

The trick is easy to miss because the announcement was dressed as a technical achievement. But look at the ladder it sits on. OpenAI already sold a Fast mode at twice the speed for twice the price; the changelog says up to 2.5x faster. Yuchen Jin, on X, summed up the pair in a line that eChai collected: “Fast: 2x faster, 2x the price. Ultrafast: 8x faster, 6x the price.” I would not press that comparison too hard, since the 8x belongs to Codex and the 6x price to the API. But the pattern holds: you pay roughly in proportion to how little you wish to wait. Nothing wrong with that. Airlines have been doing it forever, and the seats at the back still go to the same city.

Who is in a hurry? A person chatting with a model is rarely bothered by a few seconds. A person watching a coding agent work through a long job, correcting it and sending it back round, is. So is a program that calls a model dozens of times before it can answer anything at all. For them, delay is not an annoyance; it is the whole cost of the loop. Altman told the room, according to CNBC, “We think you are really going to love this and not want to go back.” Simon Willison’s live blog has him calling it worth the six-fold price. Romain Huet, on stage, drove a 3D model of the venue with Ultrafast to make the point in pictures. Anyone who has waited on a slow machine knows how quickly patience becomes the expensive part.

The rest of the day suggests OpenAI is chasing seconds everywhere. The Decoder’s caption on the new Decisions API, which asks a small model to pick from a fixed list of answers, claims about 10x faster decisions than GPT-6 Luna through the regular route (a company claim, reported by one outlet). Simon Willison recorded a member of OpenAI’s team saying that models had helped optimise the computer-use machinery for a 2x latency win. None of this is a price tier. All of it is a bet that the next thing developers will complain about is not that the models are stupid, but that they are slow.

A few honest gaps. OpenAI has not said how it achieves the speed; commenters have guessed, and I shall not join them. Ultrafast for Sol is “coming soon”, or “in the coming days”, depending on where you read it, and I have used only the 300 tokens per second that the company gave for Codex. Nor is the idea new: OpenAI’s changelog shows Ultrafast for an older model, GPT-5.6 Sol, in limited preview on 21 August, billed as up to 14x faster than standard. The tier has been in rehearsal for more than a month.

The buffet puts weights on the plates

Then comes the part that turns a pricing curiosity into a business model. Inside ChatGPT, Ultrafast in Codex and ChatGPT Work is available only on a new $500 Pro plan and on Enterprise. OpenAI’s help page says that buying extra credits on the $100 or $200 plans does not unlock it at launch, and that on the $500 plan it draws down your included usage first, then credits. The fast lane is a reason to come up a floor.

The floor above is where the sums get odd. From 30 October, the $200 Pro plan, the one people bought for Codex and Work, drops from 20 times the ChatGPT Plus allowance to 10, according to TNW, Engadget and The Decoder alike. Its weekly allowance of GPT-6 Pro messages in chat halves too, from 200 to 100. Beside it OpenAI has placed the $500 plan. The company’s DevDay recap describes it as having the “highest usage allowance at 25 times the ChatGPT Plus allowance”, plus Ultrafast. One Hacker News commenter did the sum: 2.5 times the price for 1.25 times the old allowance, before the higher burn rate of the fast lane. He was, as it happens, nearly right.

I should be careful, because the sum leans on a number that is oddly shy. The 25 times appears in OpenAI’s recap, and Simon Willison heard it on stage. It does not appear on the pricing page or the help page for the Pro tiers, and The Decoder notes that OpenAI “doesn’t say exactly how high” the allowance goes. Other users on Hacker News complained that they could not find the limits at all. So 25 times is what OpenAI said in public, and nobody has yet seen it written where a customer would look.

Grant it anyway, and the arithmetic depends entirely on where you stand. Against the old $200 plan, $500 buys 1.25 times the allowance at 2.5 times the price. Against the new $200 plan, 25 times against 10 times, it buys 2.5 times the allowance for 2.5 times the price, a perfectly straight line. The bargain was not removed from the $500 plan. It was removed from the $200 one, and the $500 plan is what the line looks like once you can see the whole thing. Add Ultrafast’s burn rate and the person who pays $500 and spends it in the fastest lane is getting through the allowance rather quickly. The plan that promises the most is the one with the most ways to spend it.

Existing $200 subscribers are cushioned, for a while. OpenAI’s help page says those whose plan was active at an “eligibility cutoff”, or in the seven days before it, keep the previous allowance through 29 October, after which it moves down. The price stays $200. The cutoff itself is not stated. TNW reports a one-time credit of $2,500 expiring at year end; The Decoder and Engadget mention only a “one-time credit”, so I would treat the figure as TNW’s alone. The old five-hour usage limit, TNW and The Decoder both say, will not be reintroduced on Pro 200, which is a small mercy dressed as a policy note. The $100 plan, according to BGR citing OpenAI’s Tibo, stays at five times Plus. The floor did not move. The middle did.

Here is the argument I would make, and it is an opinion rather than a finding. The flat monthly fee was always a mildly odd way to sell something that costs real money each time it is used. It worked like a buffet: cheap for the light eater, ruinous for the heavy one, and kept afloat because most people are neither. Whether OpenAI was subsidising the heavy eaters the company has not said. But its people have begun to describe a different meal. Sarah Friar told CNBC: “When we launched our $200 SKU, people thought we’d lost our minds… people are willing to spend more.” And The Decoder, summarising a post by Sottiaux, reports that API prices should fall over time until it makes sense for most people to buy usage as needed, and headlines the affair as OpenAI nudging users toward pay-per-use, adding that the $200 plan now gives half as many API credits per dollar. That is one publication’s summary of one post, and I offer it as such. It is consistent with the rest: if usage is to be metered honestly, the price per unit is the thing to watch, and it is falling.

Nor is OpenAI alone. Engadget, whose headline called the whole affair “OpenAI Adds $500(!) Pro Subscription, Nerfs Its Existing $200 Tier”, observed that Anthropic and Perplexity followed OpenAI’s $200 price when it first appeared, “so it’s possible other competitors will follow suit”.

None of this is sinister, and some of it is sensible. A company that sells electricity will eventually price it by the unit. What deserves a raised eyebrow is the choreography: a cheaper brain, a dearer minute, a higher rung and a lower middle, all on one day, with the size of the new top rung found in a recap page and on a stage but not on the pages a customer would check. Subscribers can forgive a rise. What they resent, as the community threads asking for grandfathering show, is not being able to read the number.

For the whole history of the internet the industry sold us instant, and sold it cheaply, because instant was the point. Now that the thinking is going cheap, it has discovered that waiting can carry a price list of its own. For anyone who has been quietly eating for two, the question for 30 October is a short one: is the meal you are used to still on the menu, and what does it cost by the gram?

Sources: