ChatGPT's new $500 tier did not raise any price — it deleted the only volume discount OpenAI sold

The announcement has two halves and only one headline

OpenAI shipped more than twenty things at DevDay. The one that moved the most money for existing customers was not a model.

The pitch for Pro 500 is clean: $500 a month, the highest usage limits OpenAI sells in ChatGPT, exclusive access to Astra Ultrafast in ChatGPT Work and Codex, and the first of the new always-on agents included. It sits above Pro 200 at $200 and Pro 100 at $100, which in turn sit above Plus at $20. A premium rung on an existing ladder — the most ordinary thing a SaaS company does.

The second half of the announcement is that Pro 200, which OpenAI had paused for new subscriptions in September, reopened with a different allowance. From 30 October 2026, included usage in ChatGPT Work and Codex drops from 20 times the Plus allowance to 10 times, at an unchanged $200. Existing subscribers hold their current limits through 29 October and receive a one-time grant of 62,500 usage credits, which OpenAI prices at $2,500 and which expires on 31 December 2026. OpenAI’s Thibault Sottiaux, converting the allowances to their API-price equivalent, put the new figure at roughly half the old one — which is what halving means, confirmed by a named source at the vendor rather than inferred from a changelog.

Those two halves are usually reported separately, as a launch and a limit change. They are one decision, and the arithmetic says so.

Divide price by allowance

Four rungs at exactly $20 per multiple of Plus, and one at half that. The outlier is not the new tier. The outlier is the plan being changed.

This is not a hostile reading — independent comparisons of the three Pro tiers landed on the same observation, noting that all three now cost about $20 per 1x of Plus usage and treating it as evidence of consistent pricing. It is consistent. A flat rate card is legible, and legibility is genuinely worth something after a year in which the list price stopped being the number that determines the bill.

But flat and fair are different claims. Every software ladder a procurement team has ever climbed gets cheaper per unit as you go up, because volume commitment is worth something to the seller. OpenAI’s subscription lineup now does the opposite of that: it costs the same per unit at $20 a month and at $500 a month. There is no bulk rate. Buying the top tier buys capacity, not efficiency.

Until 30 October there was exactly one place in the lineup where volume paid — and that is the rung being repriced.

What the $500 buys that the $200 does not

Strip out allowance, which is linear, and three things are left on Pro 500.

Ultrafast. OpenAI describes up to 300 output tokens per second in Codex, roughly eight times standard, and up to six times faster generation via the API. On the API the tier carries a published number: GPT-6 Astra Ultrafast is reported at $60 per million input and $300 per million output, six times standard Astra rates. Sol Ultrafast shipped on 8 October 2026 at $12/$60 — the same 6x multiplier, a fifth of Astra’s Ultrafast rate, and the only version of the tier a European buyer can purchase, because Astra Ultrafast supports US and global processing only while Sol’s supports EU data residency.

That price tag is the useful context for the bundle. Ultrafast is not a feature OpenAI throws in; it is a 6x multiplier that the subscription absorbs on your behalf, against the allowance you just paid for. The 25x allowance and the 8x speed draw on the same pool, in opposite directions. For a developer blocked and watching a cursor blink, that trade is obviously worth making. For anything running in the background, it is close to the worst way to spend the budget.

It is also the resolution of a thread this desk opened in August, when Ultrafast arrived as a limited preview on Cerebras silicon with no price, no SLA and no region list. Six weeks later it has a price, a consumer tier and a throughput figure — and the throughput figure is lower than the preview’s headline, because 750 tokens a second on one model in one preview is not the same product as a tier you can buy.

An included dot. The new always-on agents are bundled at Pro and above, with a significant catch for European buyers that is a separate story in its own right.

Headroom. Which is the honest reason most buyers will move, and which is precisely what the Pro 200 change manufactures demand for.

The grant is a clock, not a rebate

The 62,500-credit grant is real money and worth saying so plainly: at the $0.04 per credit implied by OpenAI’s own $2,500 valuation, it is more than a year’s worth of the allowance being removed, measured at Plus list rates. That is not a token gesture.

It is also void on 31 December 2026.

The cut begins 30 October. The credits expire 62 days later. Consuming the full grant therefore requires spending roughly $40 a day of additional usage, every day, through the end of the year — on top of whatever the seat normally does. A team that was already running at 20x Plus and now has 10x will burn through its reduced allowance and dip into credits, so some of this will be used naturally. A team that was comfortable at 12x will not come close.

Then it is January, the allowance is still halved, and the compensation is gone.

This is the same mechanism as Sol’s promotional price cut with an expiry date buried in the announcement: a dated sweetener attached to an undated change. The change outlives the sweetener by design, and the only way to see that is to write both dates on the same line.

Three repricings, one month, one unit problem

The pattern is now well enough established to name. In September, Anthropic made a 25% weekly-limit increase permanent for Claude Code, which measured against the temporary boost it replaced was a 17% cut, against a base figure that has never been published. The same month, Anthropic’s 40% cost reduction on Opus 5.5 turned out to be a change in default effort rather than a change in rate.

OpenAI’s version is more transparent than either — the multiples are published, the date is published, the grant is quantified, and a named engineer confirmed the halving. Credit where it is due.

What all three share is the unit. “20x the Plus allowance” is a ratio to a quantity OpenAI does not publish in absolute terms. You cannot convert it to tokens, to messages, or to hours without the vendor’s own conversion table, and that table is not on the pricing page. So when the ratio moves, a buyer has no independent way to price the move. The only reason the $10-versus-$20 arithmetic in this article works at all is that the ratios are quoted against a common base — and the moment OpenAI restates the base, every figure above becomes unverifiable.

That is the structural finding, and it is not specific to this week. Subscription seats for AI tools are now sold in a currency the seller mints, measures and redefines. The API side still bills in tokens, which are at least countable by the customer.

What to do before 29 October

If you hold a Pro 200 seat from before this week, you are the only customer in the lineup holding a discount, and it has nine weeks to run. Measure actual consumption against the allowance now. A seat that never exceeds 10x loses nothing on 30 October and should ignore the noise. A seat that routinely approaches 20x has a real decision, and the choices are Pro 500 at $500, or metered API usage, or less work.

If you are buying new, the ladder is flat, so pick on capability rather than value-per-unit. Pro 100 and Pro 200 differ only in allowance; Pro 500 is the only one with Ultrafast. If latency is not the constraint, the top tier is buying you headroom at the same unit price as the bottom tier.

If you are scaling a coding or agent workload, model it against the API before buying seats. Prompt caching and batch rates still discount volume on the developer platform, and GPT-6.1 Sol at $2 per million input with $0.10 cached reads changes that calculation materially — see the coding-tools roundup and the developer guide for where the seat-versus-meter line currently falls, and Codex for what the subscription allowance actually covers.

If you are comparing vendors, note that Claude and ChatGPT are now both selling subscription capacity in undisclosed base units with recent downward revisions, which means the comparison cannot be made on published numbers alone. The ChatGPT plan breakdown tracks the current tiers, and the assistant roundup now carries the per-1x arithmetic across the whole Pro ladder; EU buyers should also read why speed tiers in particular keep being the thing European customers cannot purchase.

The sentence that matters

OpenAI did not raise a price at DevDay. Every number on the rate card either stayed the same or described something new.

It discontinued the only plan in its lineup that charged less per unit for buying more, replaced it with a plan at the standard rate and the same name, and launched a tier 2.5 times the price for 1.25 times the allowance of the plan it supersedes. All of that is disclosed, dated and arithmetically checkable, which is better than most of this market manages.

It still costs a grandfathered Pro 200 subscriber half their capacity on 30 October, and the announcement that tells them so is titled after a product they are not buying.

Questions or corrections? Email Pick Right. Want the full list? See all news.