AI Shrinkflation: Your Price Held, Your Allowance Did Not, and Nobody Told You

Your AI subscription has a price you can read and an allowance you cannot. That asymmetry is the whole story. When Anthropic added weekly rate limits to Claude Pro and Max in 2025, it gave subscribers 31 days of notice. Its published commitment to developers is at least 60 days before an API model is retired. Same company, same compute, two very different duties of disclosure.

What AI shrinkflation actually is

AI shrinkflation is a reduction in what a subscription delivers while the headline price stays the same. It arrives as a smaller allowance, a change to the unit the allowance is counted in, or the removal of a model you had built a workflow around. The price tag is stable, so the increase never shows up as an increase.

The grocery analogy is useful and it is also where most coverage stops. A chocolate bar has a declared weight printed on the pack. Shrink the bar and the printed number changes, which is why a regulator can write a disclosure rule about it.

Software subscriptions sold in messages, credits, sessions or requests have no declared unit in that sense. There is a limit, it is real, and in most cases the vendor has never published what it equals in tokens. You cannot compare this month's bar to last month's when neither one has a weight on it.

The interesting claim is not that vendors are charging more. It is that the pricing model in use makes the question unanswerable by the buyer, and unanswerable questions do not get raised at renewal.

The three mechanics, ranked by how hard they are to see

Almost every case filed under shrinkflation is one of three distinct things. Separating them matters, because only one is a genuine reduction and the other two are transfers of risk.

1. The allowance cut

The most visible version. A number goes down, or a new ceiling appears where none existed. Anthropic's weekly rate limits are the clearest documented example. Two new limits reset every 7 days, layered on top of the existing 5-hour caps, effective 28 August 2025 across the $20 Pro plan and both Max tiers.

A cap you can name is a cap you can plan around, and Anthropic announced it publicly rather than shipping it silently.

2. The unit swap

Harder to see and, in my view, the more consequential of the three. The allowance stops being counted in something you can count and starts being counted in dollars of metered consumption.

Cursor is the textbook case. Before 16 June 2025 the Pro plan carried a limit of 500 requests per month, with Sonnet models costing two requests each. After the change the same $20 bought unlimited use of Tab and Auto, plus $20 of frontier model usage at API pricing. Cursor's own worked example put that at about 225 Sonnet 4 requests, 550 Gemini requests, or 650 GPT 4.1 requests.

Under the old scheme your monthly ceiling was a constant. Under the new one it moves with model choice, context length and whatever the underlying API charges next quarter. The vendor stopped absorbing that variance and you started absorbing it.

3. The model swap

The version that produces the most noise and the least measurable loss. A model you relied on disappears from the picker, or requests quietly route to a cheaper sibling once you cross a threshold.

OpenAI removed GPT-4o from ChatGPT at the GPT-5 launch in August 2025, restored it after user objection, then retired it along with several older models on 13 February 2026. The company said the vast majority of usage had already shifted, with roughly 0.1% of users still selecting GPT-4o daily. A Change.org petition still gathered more than 20,000 signatures.

The removal was defensible on usage data and it still broke real workflows for people whose prompts had been tuned against that specific model.

What actually changed, vendor by vendor

Four documented changes, all with the headline subscription price held constant.

The Copilot Pro+ tier followed the same pattern at $39 per month with $39 of credits. Users on existing annual plans stayed on legacy request-based billing until expiry.

Cursor: the request that stopped being a request

The direction of travel matters more than the arithmetic. On the old plan, 500 request units with Sonnet at two each gave you roughly 250 Sonnet turns, and 250 was 250 in January and in December.

On the new plan the same $20 buys about 225 Sonnet 4 requests. That is a modest reduction on that specific model. The larger change is that the number is now derived rather than declared, so it moves when model prices move and when your context windows grow.

Cursor's chief executive apologised on 4 July 2025 for unclear communication and offered refunds for unexpected charges between 16 June and 4 July. It confirms the users were not imagining the change.

GitHub: a credit is not a request

GitHub announced on 27 April 2026 that premium request units would be replaced by AI Credits from 1 June 2026. The new plans hold Pro at $10 per month including $10 of credits and Pro+ at $39 including $39. Usage is now calculated on tokens consumed, including input, output and cached tokens, at each model's listed API rate.

Run the old numbers next to the new ones. Copilot Pro included 300 premium requests with additional requests at $0.04. Priced at GitHub's own overage rate, that allowance was worth $12. The replacement budget is $10.

I would not lean on that comparison too hard. An overage price is not the same thing as the value of an included unit, and a short cheap request under the new scheme can cost far less than $0.04. What the comparison does establish is that the two schemes are not obviously equivalent, and GitHub has not published a conversion that would let a buyer check.

The notice gap is the finding

Here is the part I did not expect when I started checking. The same companies that give subscribers a few weeks of warning publish precise, binding notice policies for developers, and they honour them to the day.

OpenAI's API deprecation policy commits to at least 6 months for generally available models. It gives at least 3 months for specialised variants such as Codex and deep research models, and as little as 2 weeks for anything carrying preview in the name. Anthropic's documentation states that it provides at least 60 days notice before model retirement for publicly released models, with a four-stage lifecycle of active, legacy, deprecated and retired.

What 60 days actually bought

Check the arithmetic rather than the promise. Anthropic deprecated Claude Sonnet 4 and Claude Opus 4 on 14 April 2026 and retired both on 15 June 2026, which is 62 days. Claude Haiku 3 was deprecated on 19 February 2026 and retired on 20 April 2026, exactly 60. The floor is real and it is being run close to.

Now compare the subscription side. Anthropic's weekly limits were announced on 28 July 2025 and took effect on 28 August 2025, a gap of 31 days. GitHub announced the Copilot billing change on 27 April 2026 for 1 June 2026, a gap of 35 days. Cursor changed the Pro plan on 16 June 2025 and published the explanation on 4 July.

The first two rows are backed by written policies that a buyer can cite. The last three are individual decisions. That is the distinction, not a judgement about whether 31 days was enough.

Developers get a contract because their integration code breaks loudly and publicly. Subscribers get a changelog because a subscriber who runs out of allowance simply waits, and waiting is invisible to everyone except the person doing it.

Why vendors are doing this, in one paragraph of economics

A flat monthly fee against a variable marginal cost only works while heavy users stay rare. Inference is not free at the margin, and agentic tools changed the shape of the usage curve by turning one human instruction into hundreds of model calls. Anthropic said as much when it introduced weekly limits, citing users running Claude Code continuously in the background and describing itself as very constrained on compute.

The $20 price point is the second half of the explanation. It became the category anchor early, and moving off it is a public act that invites comparison. Adjusting an allowance is not. Given a choice between changing a number everyone can see and changing a number almost nobody measures, any rational pricing team changes the second one.

None of that requires bad faith. I think it is a pricing model the vendors cannot yet commit to, being sold in a format that implies they can. The unit economics behind that gap are covered in the piece on how inference costs run through AI gross margins. The same pressure is reshaping headcount-based pricing in the analysis of seat compression across SaaS.

Where this argument is weakest

Three genuine problems with everything above. The first is the one that nearly killed the post.

Limits went up in 2026, not down

Anthropic permanently doubled the Claude Code 5-hour limits for Pro, Max, Team and seat-based Enterprise plans in May 2026. It then ran a promotion giving 50% higher weekly limits, extended to 19 July 2026. Prices did not change. That is inflation running backwards, at the same vendor that introduced the weekly caps.

This is the strongest evidence against my own argument and I am not going to bury it. What it shows is that allowances are a competitive instrument, not a ratchet. When a rival ships a credible coding agent, the allowance goes up. The asymmetry of disclosure survives either way. The uplift was announced with the same advance notice as the cut: not much, and with an end date attached.

Constraint is a real explanation, not an excuse

Compute has been supply-limited for three years. A vendor that lets a small number of accounts consume a disproportionate share is degrading service for everyone else, and rationing is the ordinary answer to that. Anthropic's stated target was account sharing, reselling and continuous background use, which is not a description of your team.

If you want the scale of the constraint rather than the rhetoric, the capex piece lays out what the industry is spending to relieve it.

Nobody can produce the counterfactual

The honest hole in every shrinkflation claim, including this one, is that no independent party measures what a plan delivered last quarter against what it delivers now. Anthropic's under 5% estimate cannot be checked because the underlying usage distribution is not published. I would treat that figure as unfalsifiable rather than false, which is a different and more useful complaint.

The same applies to the widely repeated claim that models get quietly quantised or downgraded under load. I found no primary evidence for it, and I am not going to assert it on the strength of restatement volume. Six blog posts citing each other is one source.

How to detect it in your own account

You cannot audit a vendor's allowance. You can audit your own consumption, and that is enough. A change in the plan shows up as a change in what a constant task costs you.

Run a canary task. Pick one job your team does every week that is genuinely repeatable. The same repository, the same document, the same prompt, the same model, run on the same day. Log three things each time: how long it took, how much of the allowance it consumed, and whether it completed on the first attempt. Ten weeks of that data settles arguments that no amount of forum discussion will.

Record the plan terms as text, not as a memory. On the day you buy, copy the vendor's current limits page into your own notes with the date. Vendors update documentation in place. If you do not keep a dated copy, you will be arguing from recollection against a page that has silently changed.

Watch the fallback, not the failure. Modern tools rarely refuse. They downgrade. The visible signals are a switch to a mini or lightweight model after a threshold, a longer queue at peak hours, a shorter effective context, or a sudden willingness to summarise instead of read. Instruct your team to report the downgrade, because most people quietly work around it and never mention it.

Track cost per completed unit of work, not cost per seat. A seat price that holds while the number of tickets closed per seat falls is a price rise. This is the same discipline argued in the piece on where measurable AI return has actually shown up. It is the only version of the measurement you can perform without vendor cooperation.

What to write into the renewal

Everything above is diagnosis. This part changes an outcome, and it only works before you sign.

Ask for the allowance to be expressed in a unit you can count, and get the conversion in writing. If the plan is sold in credits, the vendor knows the token rate behind them, because they compute it to bill you. A supplier who will not state the conversion is telling you the conversion is expected to move.

Then ask for two clauses. The first is a notice floor on material changes to included usage, pegged to something the vendor already publishes. Both OpenAI and Anthropic have written API deprecation policies, so the ask is not novel and it is awkward to refuse. The second is a model continuity clause: advance notice when the model underneath the product is retired, replaced or re-routed. That matters because your prompts are tuned to a specific model whether or not anyone wrote that down.

Neither clause is standard on a $20 self-serve plan and you will not get them there. Both are entirely gettable on an annual team or enterprise agreement, and almost nobody asks. The same negotiating window is where the wider rationalisation work pays off, and if you are still choosing between vendors, the build versus buy comparison for coding agents covers what else to test before committing.

One more thing, which is the counterintuitive part. I would rather have a smaller allowance stated in a unit I can count than a larger one stated in a unit the vendor defines. Certainty is worth paying for, and it is currently underpriced by buyers who optimise the headline number.

Frequently asked questions

Where to start this week

Open the billing page of your largest AI subscription and find the sentence that states how much usage is included. If that sentence contains a unit you cannot convert into tokens or requests, you have found the exposure, and you have found it before renewal rather than during an outage.

Then set up the canary. One repeatable job, one fixed prompt, one fixed model, logged weekly in a spreadsheet nobody has to maintain. It costs an hour to build and it is the only instrument that will ever tell you whether the thing you are paying for got smaller. Vendors will keep adjusting allowances, in both directions, because compute economics force them to. The buyers who notice will be the ones who wrote down what normal looked like.

Related reading