ChatGPT Plans Compared: Six Tiers, One Table

Comparing ChatGPT plans in August 2026 is harder than it should be, because the tier list most search results describe is not the tier list OpenAI publishes. At the time of writing, OpenAI’s own pricing surfaces name six plans — Free, Go, Plus, Pro, Business and Enterprise — and the real difference between them sits in context windows, model access and admin controls far more than in the headline monthly price.

The timing matters. On August 6, 2026 OpenAI published a chat-side model update that changes what the bottom of the ladder is worth: GPT-5.6 Luna becomes the default model for Free and Go, and Plus and Pro get a retuned GPT-5.6 Sol plus an effort slider. That single release moves the practical gap between paying nothing and paying $20, which is exactly the comparison most people are running when they search for a plan table.

This guide reconciles both OpenAI pricing pages into one six-row reference, sets out the plan-gated context windows, maps the August 6 changes tier by tier, and names two things the vendor pages say that look wrong — one of which is a genuine anomaly on OpenAI’s own site, and one of which is an error that lives entirely in third-party aggregators. Every price below is labelled with the surface it came from.

OpenAI splits ChatGPT pricing across two pages. The individual pricing page carries Free, Go, Plus and Pro. The business pricing page carries Business and Enterprise. Neither page shows the other page’s tiers in its main comparison grid, which is most of the reason third-party summaries disagree about how many plans exist — a writer who reads one page and estimates the other ends up inventing a ladder.

The table below is the reconciliation: all six tiers in one grid, each cell taken from the page that actually publishes it, both retrieved on the same day. Where OpenAI uses its own wording — the “Best for” taglines, the “Flexible” model-access label, the “Unlimited*” footnote — that wording is reproduced rather than paraphrased, because the paraphrases are where accuracy tends to leak.

Two structural facts fall straight out of the grid. First, none of the four individual plans carries a single Security & Administration feature — SSO, admin console, SCIM, role-based access, IP allowlisting and data residency are all marked “No” down the entire individual column set. Those capabilities are not expensive add-ons on Pro; they simply do not exist below Business. Second, the price ladder is not linear: Go to Plus is a 2.5x step ($8 to $20), Plus to Pro is a 5x step ($20 to $100), and the biggest jump in the whole table buys usage headroom rather than a new feature class.

The individual ladder is cumulative in OpenAI’s own wording: each tier is described as “everything in” the tier below it, plus a named set of expansions. That structure makes the interesting question not “what do I get” but “what is the first thing I cannot get one rung down.” Read that way, the four plans separate cleanly.

The model-access row is the one most third-party tables get wrong. It is tempting to read “ChatGPT now runs GPT-5.6” and conclude that every tier gets the GPT-5.6 family. The vendor table says otherwise: Free and Go are marked “No” for Sol, Sol Pro and Terra, and Free’s model-level access point is Luna specifically. If your evaluation depends on a particular named model rather than on “whatever the default is,” Plus is the floor. We walked through what those three model names mean and how they were positioned when GPT-5.6 went to general availability, and the equivalent Plus-versus-Pro split for the previous generation is set out in the GPT-5.5 complete guide.

Notice also what Pro’s premium is not. Between Plus and Pro the security and administration column does not change at all — both are “No” on SSO, admin console, SCIM and data residency. The five-fold price step buys usage multiples, a bigger context ceiling, Sol Pro reasoning and image throughput. If what you actually need is governance rather than headroom, Pro is the wrong direction to spend and Business is the right one, at a fifth of Pro’s monthly figure per seat.

The single most under-reported line on OpenAI’s pricing page is that the context window itself is a paid feature. Most coverage treats plan tiers as message-count tiers, which is how usage caps work but not how the window works. The vendor table publishes two separate figures per plan — a GPT Instant total context window and a GPT Reasoning total context window — and both scale with price.

The arithmetic is unusually clean. Go and Plus sit at exactly twice Free’s instant window (54K against 27K). Pro sits at roughly 4.7 times Free and about 2.4 times Plus. On the reasoning side, Pro’s 400K is roughly 1.6 times the 256K that Go, Plus, Business and Enterprise all share — a 144K difference that is the widest single gap anywhere in the table.

OpenAI also publishes an approximate translation into pages of text, with a footnote that these are approximations and that the shared window also holds system instructions, memory and internal processing. On GPT Instant the maximum input runs to roughly 12 pages on Free, roughly 40 pages on Go and Plus, and roughly 250 pages on Pro — so Pro takes about 6.25 times the single-shot input of a Plus seat and roughly 20.8 times a Free one. On GPT Reasoning the ceilings are about 320 pages for Go and Plus against about 680 for Pro, a little over twice as much.

That is the number to reason about if your work is document-shaped rather than conversation-shaped. A team pasting forty-page contracts into Plus is near the published ceiling on the first message, before memory and system instructions take their share. The same team on Pro has real headroom. Message quotas are a different axis again — usage pools and how they refill are covered in our breakdown of how GPT-5.6 usage pools work across the paid tiers.

On August 6, 2026 OpenAI published a chat-side update that lands differently on each rung of the ladder — and, importantly, on a staged schedule rather than all at once. Because the rollout crosses two weeks, a plan table written on August 5 and a plan table written on August 12 describe genuinely different products at the bottom of the ladder. The table below separates what is live from what OpenAI has said is coming.

OpenAI’s stated quality claim for the update is worth reading precisely, because it is routinely reported as one number when it is two. In an internal evaluation of financial, medical and legal prompts requiring factual detail, OpenAI says responses containing at least one factual error were about 62% less common with GPT-5.6 Luna and about 68% less common with GPT-5.6 Sol than with GPT-5.5 Instant. Those are two model-specific figures from a vendor-run eval on a narrow prompt set — not a general-purpose accuracy claim, not independently audited, and not a single blended number. Coverage that prints “62% fewer errors” next to the word ChatGPT has already lost the distinction.

The Think button deserves a note of its own. For Free and Go it is the mechanism for asking for more reasoning effort without switching to a different named model — conceptually the free-tier analogue of the effort slider Plus and Pro now have, implemented as a discrete button rather than a continuous control. That is a meaningful structural shift: effort becomes a user-facing dial across the whole ladder instead of a model-selection problem. We unpack how those dials behave, and where they help or hurt, in our guide to effort dials in consumer AI, and the free-tier side of this release has its own walkthrough in the GPT-5.6 Luna free-default breakdown.

One honest wrinkle for anyone building a plan table this week: at the time of writing, the Free column on OpenAI’s pricing page already reads “Unlimited text chats with GPT-5.6 Luna,” while OpenAI’s own announcement places the unlimited-chats rollout in the following week. The marketing surface is running slightly ahead of the shipping surface. Treat the pricing-page line as forward-looking rather than as a description of what a Free account does today, and verify against the live product before you plan around it.

The business pricing page publishes one real number and one blank. Business is $20 per user per month billed annually, or $25 per user per month billed monthly, with a stated minimum of two users. Enterprise is “contact our sales team to discuss pricing,” with two documented alternative billing structures — credit-based and token-based — each linked out to its own rate-card article rather than priced on the page.

That parity is the most actionable derived fact in this whole comparison, and almost nobody states it. Two colleagues each paying for Plus are spending exactly what a two-seat Business workspace billed annually costs, and getting none of the governance. The context windows match as well — Business lists the same 54K instant and 256K reasoning windows as Plus. The upgrade path from “several people expensing Plus” to “one governed workspace” is, at annual billing, a paperwork change rather than a budget change. If a finance-shaped objection is blocking it, the objection is probably about the annual commitment rather than the rate.

Seat mechanics worth knowing before you buy

• Two seat types, one of them closing. Business runs standard ChatGPT seats at the fixed $20 to $25 per user per month, including both ChatGPT and Codex access, with the two-seat minimum. It also ran Codex-only seats, billed usage-based against workspace credits with no fixed monthly cost. As of June 24, 2026 Codex-only seats are closed to new Business workspaces — only workspaces that already had one before that date can keep adding them.
• A billing-timing change is scheduled. OpenAI has announced that starting August 19, 2026 it will charge for added paid seats immediately at the time of addition, prorated for the remaining billing period, rather than deferring the charge to the next invoice. This is a timing change, not a price change — but it does move cash out earlier for workspaces that add seats mid-cycle.
• “Team” is a retired name. ChatGPT Team was renamed to ChatGPT Business on August 29, 2025. The rename itself did not change the plan. If a comparison article you are reading still lists Team as a purchasable tier, you are reading something written before that date, or copied from something that was.

Enterprise is where the comparison stops being a table and starts being a negotiation. The exclusive column is long: SCIM, Enterprise Key Management, user analytics, domain verification, role-based access controls, data residency across ten regions, 24/7 priority support, SLAs, custom legal terms, AI-adviser access for eligible customers, invoicing and volume discounts. Both Business and Enterprise model access is labelled “Flexible” rather than a flat yes or no, which is OpenAI’s way of saying the allocation is configured per workspace and per seat. Any published “Enterprise costs X” figure is therefore someone’s quote, not a list price. The same logic applies across vendors, which is why we treat subscription, quota and token models as separate species in our comparison of the four AI pricing models.

Building a table from two vendor pages on the same day surfaces things a single-page read never would. Two of them are worth naming explicitly, because in both cases the tempting move is to quietly “fix” the number — and in both cases that would be the wrong call, for opposite reasons.

Anomaly one: Enterprise’s reasoning window is smaller than Pro’s

The individual pricing page lists Pro’s GPT Reasoning total context window at 400K. The business pricing page lists Enterprise’s at 256K. That is a 144K deficit on the plan that sits at the top of the organisational ladder, against a plan an individual can buy with a card. It is unusual enough that a careless writer would assume a typo and print 400K for Enterprise to make the ladder monotonic.

We are not doing that. Both figures are what the respective vendor pages published at the time of writing, and reporting them as-is is the only defensible choice — an inconsistency between two vendor surfaces is itself information, and silently normalising it destroys the one signal a same-day cross-read produces. If you are sizing an Enterprise deployment around long-document reasoning, this is a direct question for the sales conversation rather than something to infer from the page.

Anomaly two: the $200 tier that does not exist

The second one is not OpenAI’s error at all. A recurring claim in third-party pricing round-ups puts ChatGPT at “$8 to $200 per user per month” with a free tier and seven paid plans. Neither half survives contact with the vendor pages. Pro tops the individual ladder at “From $100/month.” Business is $20 to $25 per user. Enterprise publishes no figure at all, so it cannot be the source of a $200 claim either. There is no $200 price point anywhere on openai.com’s pricing surfaces, and there are six named tiers, not seven.

The reason this matters beyond pedantry is that pricing pages are the most-copied and least-re-verified content on the internet. A wrong ceiling propagates into budget spreadsheets, procurement approvals and board slides, and by the time anyone checks the source the figure has three citations pointing at each other. The discipline that fixes it is unglamorous: fetch the vendor page, date-stamp the retrieval, and label every number with the surface it came from. We apply the same rule to model API rates in the August 2026 AI API pricing tracker, where promotional rates and list rates get confused in exactly the same way.

OpenAI’s own “Best for” taglines are a reasonable starting point, but they describe use cases rather than constraints. The more useful question is which specific gate you hit first — named-model access, context ceiling, response-time priority, or governance. Each gate has exactly one tier where it opens.

Read the ladder forward and a pattern emerges that is likely to shape the next twelve months of consumer AI pricing. The August 6 release moved real capability downward — the default model on the free tier is now the same family paying users get, and the effort control is arriving at every rung in some form. What did not move downward is context, governance and priority. That is a deliberate structure: give away the model, charge for the window, the queue position and the audit trail. If that holds, the interesting competitive question stops being “whose model is better” and becomes “whose free tier is good enough that the paid tier has to justify itself on infrastructure.”

The practical consequence for buyers is that plan selection is becoming a capacity-planning exercise rather than a capability one. The honest way to run it is to measure before you buy: sample a fortnight of real prompts, count how often you exceed a 40-page input, count how often a slow response actually costs someone time, and count how many people are already paying individually. Those three numbers pick the tier. Teams that want that framed as a programme rather than a spreadsheet can look at how we scope AI and digital transformation engagements, which start with exactly this kind of measured baseline, or at how the same discipline applies to production output in our content engine work. And if data handling is the deciding factor rather than cost, the free tier’s specific posture is worth reading closely in our guide to what the Free tier does with your data.