OpenAI stopped selling its $200 plan because
too many people wanted it

OpenAI stopped taking $200⁰¹ a month from new customers. Not because nobody wanted to pay. Because too many did. On September 10, 2026⁰³, OpenAI paused new sign-ups and upgrades to ChatGPT Pro $200, the plan it calls Pro 20X. Existing subscribers keep it. Everyone else waits.

The stated reason was GPT-6 Astra. The model began rolling out September 3, and OpenAI head of product Thibault Sottiaux said the $200⁰¹ plan and its upgrades "put the most strain on our systems." Here is the napkin math. Plus costs $20⁰³ for one unit of usage.

That is $10 per unit of usage. Half the Plus rate. On the model that burns the most GPU time. The most expensive consumer plan OpenAI sells is also its deepest discount per unit of compute.

I do not know how many sign-ups OpenAI turned away. Nobody outside the company does. OpenAI has not published a count, a GPU number, or a reopening date. What I can show you is the pattern, and what it means for anyone pricing an AI product right now.

Yield Ladder: Sell Seats, Not Buffets

Call it the Yield Ladder. Rank every customer by one number: revenue per GPU-hour. Then sort them into rungs. When compute runs short, you close the bottom rung first and pour new capacity into the top.

Watch what OpenAI did in the same week. The bottom rung, the $200⁰¹ flat-rate buffet, got gated. The middle rungs stayed open: the $100⁰⁷ Pro tier, Plus at $20⁰³, Go at $8, and metered API access, though upgrades from Go and Plus to Pro 20X were paused. The top rung got a brand-new product: ChatGPT for Financial Services, built on Astra, with Morgan Stanley and Evercore as design partners.

That enterprise product bundles LSEG and PitchBook data into the seat. Nobody buys that at $200⁰¹ a month. Those seats carry compliance, data licensing, and a sales team, and they are priced like it. Scarce compute flowed uphill to the customers who pay the most for each token.

The rule is simple. Sell seats, not buffets. A buffet promises abundance and attracts the hungriest eaters. A seat promises access, and you know exactly how many you have. ChatGPT passed 1 billion users this year. At that scale the buffet stops being a marketing tactic. It becomes a liability on the books.

How OpenAI Routes Scarce GPUs to Morgan Stanley

You guys, let me show you exactly what is happening under the hood, because it is stupid easy once you see it. There is a hard way to price AI and an easy way. The hard way is what most builders do: pick a monthly price, promise unlimited-ish usage, and pray your costs fall faster than your users grow.

The easy way starts with one question. What does your hungriest user cost you per month in GPU-hours? Pro 20⁰³X users are, by design, the hungriest people on the platform. They run long Astra agent jobs at 20 times the Plus allowance, plus deep research and Codex, and they pay a flat $200⁰¹ for all of it.

Now layer in the supply side. SemiAnalysis reported in April 2026 that roughly half of GPU cloud providers were sold out of Nvidia H100 and H200 capacity. That is the capacity OpenAI actually needs: big reserved blocks, guaranteed.

Here is the part people miss. Spot GPU prices have fallen hard as hundreds of small providers entered the market. Reserved one-year H100 contracts went the other way, climbing from a low near $1.70²⁴ an hour in October 2025 to about $2.35 by March 2026, according to pricing trackers cited in trade coverage. That is a jump of roughly 38 percent. Cheap spot GPUs help a weekend hacker. They do nothing for a company serving a billion users at low latency.

So OpenAI did what any smart operator does when the buffet line gets too long. It closed the line. Then it opened a private dining room upstairs. Financial Services on Astra is a seat sold to banks, bundled with market data the bank already pays for, at a price per GPU-hour that dwarfs the consumer plan.

One more detail from OpenAI's own help page, and it is a monetization gem. If a Pro $200⁰¹ subscriber cancels and the subscription ends, they cannot buy the plan again until the pause lifts. Read that twice. A capacity pause just became the strongest retention mechanic OpenAI has ever shipped. Nobody on Pro 20⁰³X is churning this month.

Now the damaging admission. The contrarian read is real. OpenAI paused one SKU, not the company. The API, the $100⁰⁷ tier, Plus, and Go all stayed open, which looks a lot like tactical load management around a launch spike rather than proof that every AI business is compute-bound forever. It is unclear whether Pro 20⁰³X demand holds once the Astra novelty fades, and OpenAI has not disclosed a single utilization or margin figure that would settle it.

My read is that both things are true at once. The pause is tactical. The ladder underneath it is structural. OpenAI will reopen Pro 20⁰³X eventually, and when it does, I would bet the usage allowance, the price, or both look different.

Capacity, not demand, set all three of these prices

2031. Zoom out five years. The question is not whether GPUs get cheaper. They will. The question is who owns the reserved capacity when the next Astra-sized launch hits, and what they charge for it.

Oracle just showed you the supply side of this trade. It reported $28.5 billion in capex for its fiscal first quarter, up from $8.5 billion a year earlier, more than triple. It held fiscal 2027 capex guidance at $90 to $95 billion. Shares jumped 7 percent after hours on the news. Investors are rewarding Oracle for owning scarce inference capacity, not for selling software.

Here is the contrast pair to remember. Demand is a marketing problem. Capacity is a balance-sheet problem. Marketing problems get solved in a quarter. Balance-sheet problems take a decade, with concrete and power lines and long-term GPU contracts.

Think about the Costco hot dog. It has cost $1.50 since 1985, and Costco loses money on every one, because the annual membership is the real product. ChatGPT's free tier for a billion users is the hot dog. The Financial Services seat with LSEG data inside is the membership. The $200⁰¹ buffet sat awkwardly in between, and awkward things get cut first when the kitchen runs short.

I think the builders who win by 2031 treat compute the way airlines treat seats: reserved in advance, yield-managed by route, never sold as an all-you-can-fly pass. The asymmetric bet is on the counterposition. While OpenAI is locked into frontier-scale reserved contracts, a small builder can run a smaller model on falling spot prices and undercut on cost per task. Scarcity at the top of the market is abundance for anyone willing to serve the middle.

Hold that with some humility, though. Impermanence cuts both ways. The 38 percent rise in reserved H100 pricing could reverse the moment a new chip generation lands, and today's scarcity premium becomes tomorrow's stranded capex. Price for scarcity now, but do not sign a five-year lease on it.

Meter Your Heaviest Users This Week

Enough theory. Here is what to actually change, step by step, and you do not need a CS degree for any of it.

First, find your Pro 20⁰³X. Pull your usage logs and sort users by tokens consumed in the last 30 days. Tokens are just the chunks of text a model reads and writes, and every one costs you GPU time. Your top 5 percent of users are your buffet eaters. Write down what they cost you.

Second, build the dashboard before the feature. Copy that order. Have Codex or Claude Code scaffold a simple internal page that shows cost per user per day. Ugly is fine. A tractor beats a Ferrari here.

Third, put a ceiling on every flat-rate plan. Meta's Muse agent asks you to approve each transaction before it runs an errand. Do the same for compute. Every plan gets a monthly allowance, and heavy users see a meter, not a surprise. If your heaviest user costs more than they pay, either raise the price or cap the plan. There is no third option.

Fourth, build your upstairs dining room. Ask what data, workflow, or outcome your best customers would pay 10 times more for. OpenAI bundled LSEG and PitchBook into a bank seat. Ottermind's newly announced workspace update is a good place to prototype this: feed it the customer's source material and let it produce an editable deliverable, so you are selling the finished report, not the tokens.

Fifth, build the pause button before you need it. A waitlist toggle on your top tier takes an afternoon. OpenAI needed one on September 10 and shipped it. Yours should exist on day one.

Your first cap will be wrong. Your first cost dashboard will miss something. That is fine. Ship the meter, watch it for two weeks, and adjust. Sell seats, not buffets, and get your reps in.