CHI Special Investigation · Consumer AI Subscription
OpenAI / ChatGPTSame Price. Different Product.
You’re not buying ChatGPT.You’re buying whatever ChatGPT becomes.
ChatGPT Plus stayed $20. What that $20 bought did not stay still. Models disappeared, limits moved, features came and went, and the contract guaranteed almost none of them. The strongest case against OpenAI isn’t that it charges more for less. It’s that the thing you’re paying for can change before the subscription does.
A coherent, documented case on a narrow thesis. The broader case CHI set out to test did not survive the evidence.
The purchased bundle of models, limits and features changed often, sometimes the same day, while the price and the subscription carried on. Capacity Opacity also survives.
No Plus price increase was located. Free access expanded. Limits often rose. No consumer feature was found moving from free to paid.
Scope: OpenAI’s consumer ChatGPT product (Free, Go, Plus and Pro) — its prices, models, usage limits, features, terms and remedies. Governance, Microsoft, AI safety in general, copyright litigation and OpenAI’s executives are outside it. No numerical CHI score is issued on this page.
● 18 minute read · Research cutoff 17 September 2026Start with what failed
The case we couldn’t prove.
CHI went in testing a familiar story about AI subscriptions: that OpenAI was quietly taking value away from paying customers. Four versions of that story were tested against OpenAI’s own documents, three years of release notes and the contract. None of them held.
No increase to the US web list price of Plus was located in 43 months, or of Pro $200 in 21. Cheaper tiers were added below and between. A reported 2024 plan for a $22 Plus never happened.
Not supportedNo consumer feature was located that went from free to paid-only. The dominant direction was the reverse: paid first, then down to lower tiers and Free.
Absent as definedOne late-disclosed cut to paid Thinking budgets (Jan–Feb 2026) and one launch-day router failure. Both were acknowledged; one was partly reversed. No measurement shows a policy of serving paying users worse models.
Weakly supportedOne candidate case (GPT-4.5), transient, with incomplete primary sourcing. New Pro tiers were inserted without touching the Plus price, and Pro exclusives mostly trickled down.
Weakly supportedThe case that holds is narrower, and more specific to OpenAI: the object of the subscription itself became fluid. Between 2023 and 2026, ChatGPT moved from a stated price, named models and published limits to a capacity-governed service whose models, features and allowances could change substantially mid-subscription — while the customer’s contractual entitlement to any particular configuration stayed minimal and no change-related remedy was located.
Counter-intuitive finding
The price stayed the same.
Whatever else happened to ChatGPT, the headline prices did not climb. The ladder only grew rungs underneath and in between.
US web list price, unchanged through 17 Sep 2026. Regional and app-store prices vary.
Unchanged. New sign-ups paused from 10 Sep 2026; existing subscriptions renew.
India from Aug 2025 (₹399); worldwide from 16 Jan 2026.
From 9 Apr 2026: “5x” Plus usage, against “20x” for Pro $200.
A steady stream of capabilities that launched for paying customers later reached lower tiers and Free. This is the single strongest piece of evidence against an extraction story.
* Free deep research is a lightweight version. Free “unlimited text chats” run on GPT-5.6 Luna, subject to abuse guardrails; tools remain limited.
The hostility wasn’t in the price. It was in what the price bought.
Entitlement Drift
What $20 meant kept changing.
The record falls into three periods. The first was already mutable but legible. The second is the hostility core. In the third, OpenAI got much better at explaining change — and kept changing the bundle.
Simple, but already mutable
One price. One model. One number.A $20 price, a named model and, most of the time, a published cap. Qualitative promises of peak-time access and priority. Caps moved — down twice, up more often — and models were swapped for successors, but a customer could usually see which model was answering and what the limit was.
Directionally legible. Not stable.The hostility core
Same-day swaps. Five models gone.Same-day successor swaps in April and June. On 7 August, GPT-5 replaced five selectable models with no prior public notice located, and the product became a system: a router, a hidden mini model after limits, and from September a safety classifier that could override the chosen model per message.
Entitlement Drift: strongly supported.The churn becomes professional
Better notice. More change.Three-month legacy windows, dated retirement articles, successor mapping, automatic migration and reset times. At the same time the densest churn in the record: five model retirements, Sora, Pulse, Atlas and group chats gone, custom GPTs’ retirement announced, ads in Go, and the last numeric Plus limits withdrawn.
Entitlement Drift: supported, with improved notice.100 messages per 4 hours, then 50, then 25 per 3 hours within about a week, citing “demand and system performance”.
A published number for a named model — the high-water mark of legibility.
The replaced models left the picker the day the successors arrived.
The model behind the “o1 pro mode” Pro launched with, replaced without advance notice located.
Router, weekly Thinking cap, hidden mini fallback. Partly reversed within days (section 04).
The start of stated legacy windows for paid subscribers.
Announced 29 January. GPT-5 Instant and Thinking retired the same day at the end of a three-month legacy window announced 12 Nov 2025.
Thinking Standard, Extended and Heavy become Medium, High and Extra High effort levels.
The pricing page markets “Advanced reasoning models with GPT-6 Astra” for Plus; OpenAI describes Astra as “not yet generally available”.
No reason given. Switching for safety continues.
The defining event
August 7.
The clearest single illustration of Entitlement Drift in the record. Not because GPT-5 was a bad product, but because of what the launch did to the thing paying customers already had.
A paid subscription’s core product changed overnight.
OpenAI’s launch post said GPT-5 was “replacing GPT-4o, OpenAI o3, OpenAI o4-mini, GPT-4.1, and GPT-4.5 for signed-in users”. For Free and Plus subscribers, all five left the picker on launch day. No prior public consumer notice was located. Enterprise and Edu customers were reported to have received a 60-day window.
Teal: restored for Plus within 1–8 days after the reaction. Struck: not restored for Plus in any source located.
Not “GPT-5 was a scam”. Meaningful customer control returned only after customers reacted.
-
Five selectable models removed; a router takes over
A real-time router picks the model per request. Plus gets a weekly Thinking cap of 200 messages. “Once usage limits are reached, a mini version of each model handles remaining queries.”
-
The router fails
Sam Altman, the next day: “Yesterday, we had a sev and the autoswitcher was out of commission for a chunk of the day”; “the result was GPT-5 seemed way dumber.” Acknowledged within a day.
-
“We will let Plus users choose to continue to use 4o”
Altman also promised to double Plus rate limits as the rollout finished. GPT-4o returns behind a legacy-models setting within roughly 24–48 hours.
-
Thinking cap: 200 → 3,000 a week
A fifteen-fold increase within a week of launch, described as a test.
-
More control comes back
Auto, Fast and Thinking controls; o3 and GPT-4.1 return under additional models. Altman pledges “plenty of notice” before future removals.
-
GPT-4o retired for good
Fifteen days’ notice. Business, Enterprise and Edu kept 4o in custom GPTs until 3 April 2026; consumers did not. Whether 15 days met “plenty of notice” is a judgement call: it exceeded zero and fell below the 30 and 90 days OpenAI used later in 2026.
Capacity Opacity · When numbers turn into adjectives
The product got more powerful. The promise got less specific.
In 2023 a Plus subscriber was told what they were buying in absolute terms and, soon after, in numbers. By September 2026 the same tier is described in relative words. The comparison that matters is OpenAI against its own earlier practice, not against the rest of the industry.
Absolute promises. Published numbers.
General access to ChatGPT, even during peak timesPlus launch · Feb 2023
25 messages / 3 hoursGPT-4 in Plus · Mar 2023
50 / 3 hoursGPT-4 in Plus · Jul 2023
80 / 3 hoursGPT-4o in Plus · May 2024
100 / weeko3 in Plus · Apr 2025
unlimited access to our smartest model, OpenAI o1Pro launch · Dec 2024
Relative entitlements. Dynamic limits.
Expanded messages and uploadsPlus · pricing page
More messages with toolsGo · pricing page
Limits applyPlus footnote · pricing page
5x more usagePro · pricing page (Pro $200: “20x”)
Thinking limits depend on your plan.Model help article
Unlimited subject to abuse guardrails.Pro footnote · pricing page
On Pro $200, plus a daily allowance for GPT-5.6 Sol Pro. Pro $100 gets 50 a week. The tier that launched as “unlimited” now carries hard counts.
OpenAI’s current model help article publishes no number for ordinary Plus chat or Thinking messages. No numeric Go limit was located in any OpenAI source reviewed, at any date.
For GPT-6 Astra in Work and Codex, Plus is estimated at 5–45 messages per 5 hours. OpenAI’s own caveat: “These are not fixed message limits.”
The limits didn’t necessarily get worse. They got harder to know.
Contractual Discretion
The product lives outside the contract.
What a ChatGPT subscriber actually gets is defined on pricing pages, in help articles, in release notes and in the product itself. Those are the layers that move. The contract defines almost none of it.
- Named models
- “General access … even during peak times” (2023)
- “Priority access to new features” (2023)
- “Unlimited access” to named models (Pro, 2024)
- “Expanded”, “More”, “5x”, “20x” (2026)
OUR SERVICES ARE PROVIDED “AS IS.”
- No entitlement to any model, feature, usage limit, roster or availability level
Payments are non-refundable, except where required by law.
- 30 days’ notice of a price increase
- 30 days’ notice of Terms changes that “materially adversely impact you”
- Notice and a refund if OpenAI discontinues the Services as a whole
- A router choosing among models of different cost
- Fallback models hidden from the picker by design
- Per-message safety routing, no opt-out located
- Dynamic, un-numbered Plus and Go limits
- Retirements on 0 to 90 days’ notice
- Feature churn: Sora, Pulse, Atlas, group chats
The product lives in the help center. The contract mostly reserves the right to change it.
Remedy Asymmetry
What happens when it changes?
The company can materially change the service. The customer’s general remedy, in almost every case, is one word: cancel.
Retire it with whatever notice it chooses — 0 to 90 days observed.
Cancel. Access runs to the end of the period. No refund.
Update the Services “from time to time”. No notice period in the rest-of-world Terms.
Cancel. No change-related refund or credit was located in three years.
Adjust it. Every located reduction was announced the same day or not at all.
Cancel. OpenAI’s help centre says support does not reset usage limits.
Raise it at the next renewal.
30 days’ notice, then cancel before it takes effect.
Discontinue its Services.
Advance notice and a refund “for any prepaid, unused Services”.
OpenAI’s refund article allows discretionary refunds for an unused subscription within seven days, and for billing errors, unauthorised charges or “a documented OpenAI service or access failure”. It also says: An accidental purchase, forgetting to cancel, dissatisfaction, or a plan change does not by itself qualify outside an applicable exception.
In a product whose contents change month to month, that sentence forecloses the obvious claim.
If OpenAI shut ChatGPT down, subscribers would get notice and a refund. If a model, a feature or an app inside it disappears, the contract promises neither. The discontinuation clause was not triggered by any removal in the record, including Sora.
Refunds do exist: EU, UK and Turkey customers have a 14-day prorated right; security-related refunds were reported in 2023. The finding is bounded: no example was located of a refund or credit for a model change, feature removal or limit reduction.
Entitlement Drift · Features
Everything that disappeared.
Within roughly one subscription year, paying users lost several products they had been sold or given. They were not equally important, and most had a replacement — one OpenAI chose, not the customer.
Headline Plus and Pro inclusion from Dec 2024. Discontinuation announced 24 Mar; app and web ended 26 Apr 2026. API end scheduled for 24 Sep 2026, after the cutoff.
About 33 days
None as a capability. Purchased credits redirected to Codex.
Credit conversion; refund questions routed to the general policy. No pro-rata refund documented.
Pro-only from Sep 2025; a Plus expansion was planned but never documented. Sunset from 17 Jun 2026.
14 days
Scheduled tasks (Plus and above)
None located
Launched Oct 2025. Shutdown announced 9 Jul; closed 9 Aug 2026.
31 days
Desktop-app browser and Chrome extension
None located
Introduced Nov 2025; retired 9 Jul 2026. Existing chats became read-only.
No advance notice located
None named
Not a paid entitlement
Removed from GPT-5.5 Instant and Thinking on 28 May 2026, per release notes.
Not established
Writing and code blocks
None located
Retirement announced 11 Sep 2026: creation scheduled to end 25 Sep; GPTs scheduled to stop running 11 Dec 2026. The pricing page still lists “custom GPTs” for Plus.
About 2 weeks / about 3 months
Migration to plugins; custom actions do not transfer
None located
The counter-intuitive section
The limits didn’t necessarily shrink. They became harder to define.
Measured over three and a half years, the direction of ChatGPT’s allowances was mostly upward. The downward changes are few, dated and often contested. What moved in the wrong direction was predictability.
Most of the movement was generous
Few, dated, and mostly without notice
* Subject to abuse guardrails; tools keep separate, unnumbered limits. No advance notice was located before any reduction, and no pro-rata credit for one.
When limits are reached, a smaller model takes over. OpenAI’s April 2026 release note: Because it serves as a fallback, it won’t appear in the model picker.
Deep research, Codex and Pro allowances have counters or reset times. No pre-emptive meter for ordinary Plus chat or Thinking messages was located.
Since September 2025 a sensitive message can be routed to a different model “regardless of which model a person first selected”. Pre-announced; no opt-out located. A substitution, not a downgrade.
Credits from Nov 2025 (non-refundable, 12-month expiry) and paid instant resets for Work and Codex. Metered overflow — not Forced Monetization, because nothing free moved behind it.
A secondary finding
Ads arrived. Not where you might think.
Advertising is a real change to what two tiers deliver, but it is not a clean extraction story. Plus and Pro remain ad-free.
Became ad-supported at the same price
- Launches in India. No advertising disclosure in the launch coverage reviewed.
- Rolls out worldwide; the same day, OpenAI announces an ads test for Free and Go in the US.
- Ads live in the US, labelled “Sponsored”, below the answer. Price unchanged.
- Pricing page:
This plan may include ads.
There is no ad-free Go option; avoiding ads means switching to Free’s Ads-Free mode or upgrading.
Gained ads — and a great deal more
- Ads arrive, with an “Ads-Free” mode in exchange for lower usage limits and reduced tools. The size of the reduction is unpublished.
- Free gains unlimited everyday text chat on GPT-5.6 Luna.
- Every branch of the free user’s choice is at least as good as the early-2025 free tier on text messages. Not a clean extraction case.
Ad-free, as announced
- OpenAI: Plus, Pro, Business and Enterprise “will remain ad-free”.
- OpenAI says ads do not influence answers and advertisers do not see conversations. No independent test either way exists in the record.
- More than 40 countries by 31 Aug 2026; contextual-only in the EEA.
Privacy is not the thesis of this investigation, and these findings do not carry it. They belong here only because they are another part of the product that changed underneath the price.
Mandatory counter-case
The evidence against our own case.
This investigation only works if the reader can see what it rejected and why. Each item below changed a verdict.
The price did not move
Plus $20 since February 2023; Pro $200 since December 2024. Cheaper tiers were added rather than raising existing ones. The contract commits to 30 days’ notice before any increase.
Effect: price hostility absent; Forced Monetization absent.Paid features flowed downward
Search, reasoning, images, memory, Projects, Codex, deep research (lightweight), scheduled tasks and, in August 2026, unlimited text chat all reached Free.
Effect: no free-to-paid movement located.Allowances mostly rose
Flagship caps, Thinking after the GPT-5 backlash, deep research and memory all increased. Reductions are confined to a handful of dated instances.
Effect: the issue is mutability, not shrinkage.Disclosure improved materially
Legacy windows, retirement notices of 15 to 90 days, successor mapping, automatic migration, release notes, usage counters, reset times, a published explanation of rate limits, ads principles published before the first test.
Effect: Entitlement Drift reduced to supported for Nov 2025 – Sep 2026.The constraints were real, and stated
“Our GPUs are melting” (Mar 2025). GPT-5 Pro withheld from Plus: “we do not have the compute” (Aug 2025). On 10 Sep 2026 OpenAI paused new $200 Pro sign-ups rather than dilute existing subscribers. Its help page gives no reason; Fortune attributes the pause to GPT-6 Astra demand.
Effect: the capacity-management thesis is partly supported.OpenAI kept admitting and undoing
GPT-4o restored within days; Thinking caps raised fifteen-fold; a sycophantic update rolled back in three days; the router failure acknowledged the next day; the 2026 Thinking-budget cut disclosed and partly reversed — about 25 days after it began.
Effect: inconsistent with deliberate degradation as policy.The residue on which the PARTIAL verdict rests. None of it depends on intent.
- The 2023 “even during peak times” and 2024 “unlimited access” promises were absolute, never contractual, and were later withdrawn or re-described.
- The GPT-5 launch removed products customers were paying for, with no prior public notice located. The fast reversal proves they were missed.
- Numeric Plus limits published from 2023 to 2025 are no longer published, while Pro’s are. Peer vagueness does not cure a within-company regression.
- The router’s decision rules were never published in a form that allowed prediction; the fallback model is hidden by design.
- A paid Thinking level was reduced for about 25 days and disclosed after the fact.
- The refund policy excludes “a plan change” in a product whose plan content changes monthly; no change-related remedy was located.
Final synthesis
Partial.
Eight mechanisms were tested. Two survive as findings specific to OpenAI, two survive as structures the industry shares, three do not survive, and one is documented but is not hostility at all.
Models, caps and features changed often inside a constant-price subscription; five models removed with no prior public notice located on 7 Aug 2025.
Numeric Plus and Go limits withdrawn while Pro stays numbered; hidden fallback; estimate bands. Relative to OpenAI’s own practice, not to peers.
OpenAI can change the service; the customer can cancel. No change-related refund or credit located. Made consequential by frequency of change.
The contract defines essentially no entitlement to a model, feature, limit or availability level. Peers draft the same way.
The late-disclosed Jan–Feb 2026 Thinking cuts and the Aug 2025 router outage. No policy; no independent measurement of the app.
GPT-4.5 is the one candidate: transient, and primary sourcing incomplete. Not a pattern.
No consumer feature located moving from free to paid-only. Credits, paid resets and ads are different mechanisms and are not counted here.
Higher tiers get stronger models, more reasoning, more context and more usage. Explicit, disclosed segmentation; only its unpredictability counts, under Capacity Opacity.
OpenAI did not steadily charge ChatGPT users more for less. In many respects it did the opposite: the $20 price stayed still, free access expanded, limits often rose and newer models replaced older ones.
But the subscription became harder to define. The customer who once bought a named model with a visible limit increasingly buys a moving bundle of models, routers, fallbacks, effort levels and features whose composition OpenAI can change while the subscription continues — with notice that has improved but is not promised, and with no remedy located short of leaving.
The strongest finding is therefore not degradation. It is impermanence.
ChatGPT improved.The entitlement dissolved.
Method & limits
How this was established.
Evidence base and cutoff
The evidentiary cutoff is 17 September 2026, 23:59 UTC. The investigation synthesised ten research streams (pricing, models and routing, usage limits, quality, entitlement changes, advertising, data and memory, enforcement and refunds, terms, and a dedicated counter-evidence stream) under a common brief that tested a competing thesis — legitimate, disclosed capacity management — with equal weight. The record rests on OpenAI’s own terms, policies, help articles, release notes, pricing pages, status write-ups and posts, executive statements as reported verbatim by credible press, and peer terms from Google, Anthropic and Microsoft. It comprises a 236-row source ledger, a 166-claim audit, a load-bearing verification pass and three adversarial reviews (factual integrity, thesis challenge, synthesis integrity).
Every finding compares three layers: marketing (what buyers were told), contract (what OpenAI promised or reserved) and product (what was delivered). The gaps between layers are the object of the investigation.
Live wording re-checked for publication
The current wording quoted on this page from OpenAI’s pricing page, refund article, model help article and Terms of Use was re-checked against the live pages on 18 September 2026. OpenAI’s help articles change frequently, sometimes more than once a day; a fetch of one URL variant of the model help article during that check returned an older version that still listed a Plus message figure. The current, canonical version publishes no numeric ordinary Plus chat or Thinking cap. Quotations should be re-checked again at any republication.
Limits of the record
Web archives were unreachable throughout the research, so 2023–2025 page wording is reconstructed from dated OpenAI posts, help-centre change notes and contemporaneous press that quoted the pages. The 2023 Terms of Use could not be retrieved; this page makes no claim about when arbitration or any notice clause first appeared. Because 2026 events are documented from live pages and earlier ones largely from quotations, the density of change in the most recent period is partly an artefact of what could be retrieved. “No example located” always means a bounded search found none, never that none exists.
Scheduled events after the cutoff
The end of the Sora API (scheduled 24 Sep 2026), the end of custom-GPT creation (scheduled 25 Sep 2026) and the date custom GPTs are scheduled to stop running (11 Dec 2026) were announced before the cutoff but had not happened by it. They are described on this page as scheduled, not executed.
Why there is no CHI score
This is a bounded Special Investigation of one consumer subscription relationship, not a full company assessment. No locked score exists for OpenAI, and CHI does not derive one from mechanism verdicts. The page reports qualitative verdicts on a five-point scale (absent, weakly supported, supported, strongly supported, unresolved) instead.
What this page does not claim
- That OpenAI deliberately degrades, “nerfs” or serves quantized models to paying customers.
- That OpenAI systematically removes value from Plus to sell Pro, or that GPT-4.5 proves it.
- That OpenAI repeatedly raised subscription prices, or that ChatGPT’s quotas have generally fallen.
- That any consumer feature moved from free to paid.
- That OpenAI never gives notice, or never gives refunds.
- That every model retirement happened without notice.
- That Go was promised to remain ad-free, that ads influence answers, or that advertisers receive chats.
- That Pro is exempt from the 5-hour usage window.
- That any regulator has found OpenAI’s subscription practices unfair.
- That OpenAI gives less notice than all its competitors. No systematic cross-vendor comparison of notice periods was done.
What remains unresolved
- Whether consumer Pro was exempt from the 5-hour window at the cutoff: reported in August 2026 as exempt “for the coming months”, but OpenAI’s September help article lists Pro under 5-hour limits.
- Primary confirmation that Plus had GPT-4.5 from March 2025 — the basis of the one Tier Escalation candidate.
- The size and direction of the April 2026 Plus Codex “rebalancing”.
- Whether Go’s launch help article reserved the right to show ads.
- Whether a Plus Thinking cap still applies internally, and its size.
- The unpublished limits of Free’s Ads-Free mode, and whether existing Go subscribers were notified before ads arrived.
- The exact wording of the 2023 Terms of Use.
Continue