OpenAI Reveals ChatGPT Pro Max Subscription at $500 per Month
Researcher Tibor Blaho discovered a subscription configuration for ChatGPT Pro Max, priced at $500 per month, indicating that it offers "the fastest Work and Codex" compared to the existing Pro version. The official pricing page still lists Free, Go, Plus, Pro, Business, and Enterprise, with the company yet to confirm a launch date. This comes about five days before the San Francisco DevDay on September 29.
The current self-service top-tier Pro is priced at $200 per month, with usage approximately 20 times that of Plus. New accounts were paused on September 10 due to demand for GPT-6 Astra; a mid-tier option was added on April 9 at $100 per month, with usage about five times that of Plus. The $500 price is 2.5 times the $200 tier, but specific multipliers, model lists, or whether existing users can upgrade with one click have not been detailed in the code. Screenshots from taxed regions have shown a $600 display.
Work is aimed at long tasks, while Codex can run programming jobs continuously for hours to days, making inference latency more valuable than single Q&A. OpenAI has utilized Cerebras for low-latency paths, including GPT-5.6 Sol Ultrafast, and the external community associates "fastest" with additional wafer-level inference capacity, though neither party has confirmed this connection. Anthropic's most expensive Claude Max is still significantly below the $500 tier.
The tiering reflects a shortage of computing power being translated into pricing. The closure of the $200 tier indicates that supply cannot meet the demand from willing-to-pay agents, while the introduction of the $500 tier changes the queue into a bidding system. Long tasks and overnight Codex consume continuous tokens, not chat counts, meaning the unit price must cover peak cluster costs. Unreleased strings cannot serve as contracts but indicate that the product tree has reserved space for higher tiers.
In market mechanics, what is sold is faster agent hours, while what is bought is inference slots ahead of $200 users. Demand comes from developers treating models as labor, while supply is constrained by GPU and Cerebras-like capacities. The beneficiaries are OpenAI, which can convert latency differences into subscription differences, while heavy users already queued at the $200 tier and competitors must adjust their pricing. Funds flow from monthly fees into inference bills, determining whether to reopen the $200 tier.
Source: Public Information
ABAB AI Insight
ChatGPT has segmented the same model into Free, $20, $100, and $200 tiers, and is now preparing to introduce a $500 tier. Each tier corresponds not to a list of features, but to throttling multipliers and queue priorities. Astra has filled the $200 tier, indicating that marginal users of the flagship model are willing to pay more than $200. Pro Max uses "fastest" rather than "most features" as a selling point, acknowledging that the bottleneck is speed, not skill tree.
The capital path is to cover the cash gap between training and inference through subscriptions. APIs are sold by token, while ChatGPT is sold by seat; long-range agents bring seats closer to monthly compute contracts. If Cerebras enters the highest tier, it would mean productizing dedicated inference hardware into a subscription SKU, rather than just writing it on the system card. DevDay is a window for turning source code strings into a price list.
Similar tiering can be seen in AWS instance families, Bloomberg terminals, and Adobe Creative Cloud: first capturing users with mid-tier offerings, then opening a separate pipeline for those willing to pay for reduced latency at the top tier. The industry is at a stage of bifurcation between consumer chat and application agents. Whoever first marks "overnight coding tasks" with stable latency can charge $500.
Structural judgments indicate a transfer of pricing power. The pricing unit for model capability shifts from "can it be asked" to "can multiple tasks be completed continuously." The mechanism is: when supply is tight, raising prices clears the queue faster than expanding capacity; the tiers in the source code serve as ration coupons before the arrival of expansion commitments.
ABAB News · Cognitive Laws
Seats in the queue will definitely be more expensive at the next tier
Long tasks buy latency, not just another model name
The price list in the source code is a ration coupon for insufficient supply