ChatGPT Pro Max: The $500 Price Leaked, the Limits Didn't

What leaked, and what didn't

On September 24, 2026, Tibor Blaho posted on X: "OpenAI is preparing a new higher-tier ChatGPT Pro plan". TestingCatalog wrote it up the same day as ChatGPT Pro Max at $500 a month, found in unreleased ChatGPT subscription references. The next day BleepingComputer reported that "the unannounced plan has started appearing in ChatGPT's subscription interface alongside Plus and existing Pro options."

Two things to keep straight. The name: if you've seen "Codex Pro Max", the reports call it ChatGPT Pro Max, and Codex appears in the card's one new line, "Fastest Work and Codex". The status: OpenAI has published nothing about it. Its Codex pricing page, as of September 26, covers Free, Go, Plus, Pro, Business, Edu and Enterprise, and no Pro Max. Pro Max details not marked as my reading come from the plan card and reporters' reading of it. Current-plan facts come from OpenAI's own pages unless another source is named.

What the card says

Sources: TestingCatalog and BleepingComputer. The two prices come from BleepingComputer. TestingCatalog's video shows $600. The "Also listed" line is BleepingComputer's reading.

Work is ChatGPT Work, OpenAI's product for longer tasks, and Codex is its coding agent. TestingCatalog reads the pairing as aimed at developers, researchers and professional users running long agentic workloads, where a Codex job can run for hours or even days and speed is worth more. BleepingComputer notes that power users have been asking for more usage than the $200 plan offers. It also flags a catch: the $200 plan already promises early access, so it isn't clear whether that becomes exclusive to Pro Max.

What OpenAI already publishes

Pro Max has no official page. The plans around it do, and they frame every question the card leaves open.

• Work and Codex are one allowance. The pricing page says it plainly: "ChatGPT Work and Codex share usage." Same pricing, same credits, same limits. A faster lane for "Work and Codex" is a faster lane on one meter.
• Pro comes in two sizes. From $100 a month, with 5x or 20x more Codex usage than Plus. Fortune puts the 20x tier at $200.
• Limits are estimates per five hours. OpenAI publishes local-message estimates per five-hour period, by model and plan, says they are not fixed limits, and adds that weekly limits may also apply.
• Speed already costs allowance. Per the speed page, Fast mode uses credits at 2.5 times the Standard rate on GPT-6 Astra, Sol and Luna where available. For GPT-5.6 it lists a 1.5x speed gain for that 2.5x. For the GPT-6 models it lists no speed figure at all. The pricing page adds that speed configurations "also use included limits faster."
• Priority already exists. Enterprise and Edu plans list "Priority request processing".

Current-plan prices and ranges come from OpenAI's pricing page, which calls the ranges estimates. The $200 price, and the pause on its new sign-ups and upgrades since September 10, come from BleepingComputer and Fortune. Pro Max ($600 with VAT in one version) is unannounced. BleepingComputer reported it in the subscription interface. TestingCatalog reported it in unreleased subscription references.

The empty cell is the story. At $200, OpenAI estimates 100 to 900 Astra messages every five hours. At $500 the allowance is unpublished. $500 is 2.5 times $200, and Fast mode consumes credits at 2.5 times the Standard rate. Design or coincidence, nobody has said.

The Cerebras guess

Both reports reach for the same explanation of "fastest". TestingCatalog notes that OpenAI already uses Cerebras for its latency-focused inference path, including GPT-5.6 Sol Ultrafast, which makes extra Cerebras capacity plausible. BleepingComputer adds the history: early this year Codex-Spark launched on Cerebras at more than 1,000 tokens per second, and OpenAI later showed GPT-5.6 Sol at up to 750 output tokens per second in its Cerebras-powered Ultrafast mode.

Then both stop in the same place. In BleepingComputer's words, "neither OpenAI nor Cerebras has confirmed that connection." Those speed figures belong to other launches, and copying either onto Pro Max would invent a spec. "Fastest" could mean Fast mode, priority processing, a Cerebras path, or a mix. The card doesn't say which.

Why $500, and why now

Fortune reported on September 11 that OpenAI had paused new sign-ups and upgrades to its $200 plan because demand for GPT-6 Astra outran capacity. Thibault Sottiaux, OpenAI's head of product, wrote on X that the $200 plan puts the most strain on OpenAI's systems. He said, "We wanted to take the smallest step that allows us to continue giving the broadest access possible," adding that OpenAI is "working on adding more capacity" and that existing Pro accounts are unaffected. Fortune also reports OpenAI saying Astra can burn through usage allowances faster than GPT-5.6 Sol.

My reading, and only mine: a company that has paused new sign-ups and upgrades to its $200 tier for lack of capacity is preparing a $500 tier whose headline is speed. If it ships as described, one reading is rationing by price; another is a premium lane on new capacity. The number that decides whether it is worth paying is the one missing from the card.

TestingCatalog puts the price in context: in its current form, Pro Max would sit far above most individual AI subscriptions, including $200 plans from Anthropic and Cursor and Google's roughly $250 AI Ultra tier. It also asks the right question: "The main question is whether OpenAI plans to justify the $500 price primarily through speed, larger usage allowances, longer-running Work sessions, or a combination of all three."

Who a fast lane is for

This part is opinion. Speed pays when a person is waiting. An interactive Codex session, a review you're sitting through, a terminal you're watching: faster output means shorter waits, and the wait is the expensive part. An unattended loop is different. A headless run doesn't care whether a task takes eight minutes or four. It cares whether the allowance lasts through the night.

Today's rate card makes the trade explicit. Fast mode on Astra consumes credits at 2.5 times the Standard rate where available, so on a fixed allowance a faster lane mostly reaches the limit sooner. As described so far, Pro Max is a clear offer for someone blocked on latency and a weak one for someone blocked on limits, unless the limits rise too. Nobody has confirmed that they do.

For a Codex user, the line that matters most is whether "most capable frontier models" means GPT-6 Astra at ultra, and how much of it fits in a window. Your Agents Eat Your Subscription covers why subscription-authenticated agents can draw on the same allowance as interactive work.

Questions to answer before paying $500

• Limits in writing. Pro Max's row in OpenAI's own estimate table: Astra messages per five hours and any weekly limit, next to the $200 plan's 100 to 900.
• Models and effort. Whether Codex gets GPT-6 Astra at ultra, and whether subagents get the same.
• What "fastest" is. Fast mode, priority processing or a new serving path, with a number: tokens per second or time to first token.
• What speed costs. Whether the fast lane still spends credits at 2.5 times the Standard rate, or comes with its own budget.
• Where it applies. Codex CLI, the IDE extension and cloud chats, or only inside ChatGPT.
• The $200 plan. Whether it reopens, and whether its early access becomes exclusive to Pro Max.

When we'll know

OpenAI DevDay is Tuesday, September 29, 2026, at Fort Mason in San Francisco. The opening keynote starts at 10:00 a.m. Pacific and is livestreamed. TestingCatalog expects the event to cover APIs, developer tooling, new subscription tiers for the OpenAI Platform and new products. That is its expectation, and Platform tiers and ChatGPT plans are different products. OpenAI hasn't said Pro Max will be there, and the day can pass without a word about it.

What this does not show

• That the plan will ship. Unreleased plan references can change or disappear.
• The final price. $500 on one page, $600 with VAT on another. The real number is OpenAI's to publish.
• That every screenshot is real. Decrypt, which also covered the leak, says it isn't confirmed whether one screenshot of the interface is a leak or a render.
• Limits, credits or named models. None are on the card.
• Cerebras. Neither company has confirmed it.
• Speed. Nothing has been measured, by me or in any report.