ChatGPT Pro Max: What the $500 Report Says, and What OpenAI Hasn't Confirmed

On September 24, 2026, the leak-tracking account @testingcatalog — operated by Tibor Blaho — published what it says it found inside unreleased ChatGPT subscription references: a plan named ChatGPT Pro Max, priced at $500 a month. The $600 figure that has been circulating alongside it is the number shown in the account's own video, which it attributes to VAT. That is the whole evidence base. There is no announcement, no launch date, no model ID, and no published allowance attached to any of it. What makes the report worth reading anyway is the tier it would sit above — the one the vendor stopped selling partway through September — and the reason it stopped selling it is the model at the centre of the story: demand for GPT-6 Astra, the flagship that reached general availability on September 5 and that ChatGPT's own usage caps display as GPT-6 Pro. The cheaper API tiers, GPT-6 Sol at $2 / $10 and GPT-6 Luna at $0.10 / $0.50 per million tokens, are what you buy when a monthly seat is the wrong shape for the work. Pro Max would be the seat. DevDay is September 29, four days out.

What the report actually contains

TestingCatalog's write-up is short, and the useful thing to do with it is separate the three claims it makes from the ones it explicitly declines to make.

What it states: the plan appeared in unreleased ChatGPT subscription references; it would cost $500 per month; and its description differs from the current Pro tier "mainly through access to the 'Fastest Work and Codex'".

What it adds without confirming: that greater usage allowances might come with it — the article's own words are that this "is not confirmed yet". Our briefing for this piece puts the same caveat differently: limits may still rise before anything ships. Both formulations point at the same gap. Open​AI has not published a single number for this tier, so nobody outside the company knows whether $500 buys more messages, longer agent runs, a faster serving path, or merely a higher ceiling that most buyers never reach.

What it does not claim: no model ID, no release date, no capacity commitment, and no confirmation from Open​AI. Every outlet that has picked the story up in the last day — and there are many, from Crypto Briefing to Futubull to Startup Fortune — frames it as reported or leaked rather than announced. That framing is correct and worth preserving, because the difference between "Open​AI is preparing" and "Open​AI has launched" is the difference between a plan you can buy and a string in somebody's developer tools.

The one piece of the description that carries real information is "Work and Codex". Those are the agent buckets in the ChatGPT product, and they are metered separately from Chat: Chat, ChatGPT Work, and Codex hold distinct allowances, with Work and Codex sharing a single pool between them, and an Open​AI API key sitting outside all three. A plan whose selling line is faster access to Work and Codex is not a plan about chatting more. It is a plan about running long agentic jobs, which is the workload where a frontier model's cost per hour is hardest to predict and easiest to blow through.

The capacity story underneath it

A tier above a tier that is not for sale is not a contradiction, and reading the two together is what makes the report more than a price tag.

In mid-September — outlets reporting the pause cluster around September 11 — Open​AI stopped taking new sign-ups and upgrades to the $200 Pro plan, citing demand for GPT-6 Astra. The company's head of core product and platform was quoted saying he had "not seen anything like it", and coverage noted that the pause came even after Open​AI had expanded available compute roughly 9.5-fold since 2023. Our own coverage recorded the same thing from a different direction: as of late September, Open​AI still was not selling new $200 Pro seats, and the $100 Pro tier introduced around the same time was the route on offer instead.

So the sequence is: demand for the flagship outran capacity, the top tier was closed to new buyers, a cheaper tier was opened beneath it, and now a more expensive tier has been reported above it. Every one of those moves is a rationing decision. Closing a tier rations by queue; opening a cheaper one rations by allowance; adding an expensive one rations by price. If Pro Max is real, the interesting question is not whether $500 is a lot — it plainly is — but which of those three levers Open​AI is choosing to pull at the top, and whether the $200 tier reopens before Pro Max appears or simply becomes the tier that no longer sits at the top.

Cerebras, and what "fastest" would have to mean

TestingCatalog's article notes that "context also points to Cerebras infrastructure", and adds that Open​AI has not confirmed any direct connection between Cerebras and Pro Max. Our briefing says the Ultra Fast path is likely Cerebras-backed. Treat both as inference, because that is what they are: a plausible reading of an unconfirmed description line, not a capacity deal anybody has signed in public.

There is a real precedent to reason from, though. Open​AI already leans on Cerebras for a latency-focused inference path — the GPT-5.6 Sol Ultrafast preview, which pushes Sol to as much as 750 output tokens per second on Cerebras hardware, is a limited preview for a small set of customers rather than a generally available tier. If "Fastest Work and Codex" means anything technical, that is the shape it would take: the same weights served on faster silicon, sold to a small number of very high-paying accounts. That reading has a consequence worth holding on to. A speed tier is a serving arrangement, not a new model. Nothing about GPT-6 Astra's weights, context window, or benchmark scores would change on a Pro Max seat, so nobody should expect a Pro Max-only model to appear on the API.

What $500 a month buys, priced two ways

This is where the leak becomes a decision rather than a rumour, because a reader weighing $500 a month has a second option that is priced in public.

A ChatGPT seat is metered in messages and agent runs, and Open​AI does not publish a token equivalent for any of it. That gap is the honest starting point: you cannot convert a Pro Max allowance into tokens, because the allowance has not been published. What you can do is price the alternative. On Open​AI's list rates, $500 spent as metered API usage buys 50 million input tokens or 10 million output tokens of GPT-6 Astra at $10 / $50 per million — and 250 million input or 50 million output tokens of GPT-6 Sol at $2 / $10. Those are Open​AI's own list prices, and GPT-6 Astra's $10 / $50 doubles to $20 / $75 once a request crosses the 272,000-token line, so a long-context agent run costs more per token than the headline suggests.

The practical split is not subscription versus API; it is which workload sits on which meter. Conversations, one-off questions, and anything where you are paying for convenience belong on a seat. Overnight agent runs, batch jobs, and anything you can measure in tokens belong on a metered key, where the bill tracks the work instead of a ceiling you cannot see.

OrcaRouter is the metered side of that split. We route GPT-6 Astra at Open​AI's list price, alongside GPT-6 Sol at $2 / $10 and GPT-6 Luna at $0.10 / $0.50, with 0% markup — the provider's list price is passed through, so a vendor price change is live on our side the same day rather than at the next billing cycle. One key reaches 200+ models, and automatic failover means a model you have not committed to can sit behind a route instead of in your production path: if GPT-6 Astra is the wrong bet for a job, the fallback is a routing change, not a new contract. To be explicit about what we do not sell — ChatGPT Pro Max is a subscription product inside ChatGPT, not an API model, and we do not resell ChatGPT seats or claim any access to them. We meter tokens on the models.

What would make this checkable before September 29

Four days is long enough for the leak to resolve one way or the other, and the surfaces that would settle it are specific.

The plan picker in the ChatGPT web client is the one to watch, because that is where the references were reportedly seen in the first place. A tier that moves from a string to a selectable row in that list is a tier that exists. The pricing page is the second surface: if Open​AI puts a number on it, that number supersedes $500, and the VAT gap stops being a footnote. The third is the help centre's plan comparison, which is where Open​AI publishes the message counts it is willing to stand behind — an empty cell there means the allowance is still unpublished regardless of what the marketing page says. And the fourth is DevDay itself on September 29, which is where a new top tier would be least surprising and where Open​AI would also have to say whether the $200 tier reopens.

What none of that would change: the API. If Pro Max ships, GPT-6 Astra keeps its list price, keeps its context window, and stays reachable by key at the same rate as before. A subscription tier is a packaging decision about how Open​AI sells access to a model that already exists and is already priced in public.

So the decision for a reader today splits cleanly. If your work is conversations and short agent runs on a seat, there is nothing to do yet: the tier is unannounced and unpriced in any way you can act on, and the pie to wait for is DevDay. If your work is measured in tokens, the $500 question already has an answer that does not depend on the leak at all — the same flagship is available by the token today, at list price, with the ceiling set by your budget rather than by an allowance Open​AI has not described.