Anthropic’s usage gimmicks are becoming a trust problem

In late May, I helped a client set up Claude to work across her inbox, files and projects. She asked what plan she should buy.

I told her I paid US$200 a month for Max, used the best model all day and never hit the limit. She upgraded on the spot. Her view was simple: this costs less than employing someone, and she didn’t want to think about usage every time she gave it a job.

Two months later, I was comparing notes with another business owner. We were both hitting limits. He had already split his automations into smaller jobs to conserve usage. I suggested he use Sonnet for more of the routine work instead of leaving everything on the best model.

That is a big shift in eight weeks. I went from telling customers that the top plan gave them enough room to get on with their work, to helping them design around invisible ceilings.

Claude hasn’t become less useful. It has become much more useful. That is why this matters.

Six months of moving goalposts

Any one of Anthropic’s changes can be explained. New models cost more to run. Promotions end. Demand is hard to forecast. Products change while they mature.

Taken together, though, the pattern looks less like a stable product and more like live price discovery being run on paying customers.

Here is what Anthropic communicated to my account over the last six months.

5 February: Anthropic launched Claude Opus 4.6 and agent teams. The launch email included US$50 of extra usage for existing Pro and Max subscribers. It also warned that running several agents could increase token usage proportionately.

4 April: Anthropic offered a one-time extra-usage credit equal to my monthly subscription price. It also introduced pre-purchased usage bundles at discounts of up to 30%. In the same email, it said third-party harnesses connected to a Claude account would begin drawing from paid extra usage instead of the subscription.

That is an unusual combination: change what the subscription covers, give the customer a credit to soften the change, then sell discounted capacity once the credit runs out.

13 May: Anthropic announced a temporary 50% increase to weekly Claude Code rate limits through 13 July.

May and June: Anthropic told users that Claude Agent SDK, claude -p and third-party apps built on the SDK would stop drawing from subscription limits on 15 June. Max 20x users would instead receive a separate US$200 monthly SDK credit. Anything beyond it would move to paid usage credits at standard API rates.

On the day the change was due to take effect, Anthropic paused it. Its current support page preserves the abandoned plan for reference, including the credit table, while stating that the change is no longer taking effect.

2 July: Anthropic made Claude Fable 5 available to subscribers. It could consume up to half of a weekly plan limit at no extra cost, but only through 7 July. After that, continued access was meant to use paid credits.

8 July: The included period was extended through 12 July.

12 July: It was extended again through 19 July. The temporary 50% increase to Claude Code rate limits was extended at the same time. Subscribers were again told that Fable would move to paid credits after the promotion.

20 July: The position changed again. Fable became a standard part of Max and premium plans, still capped at half of the shared weekly allowance. Pro and standard Team users moved to pay-as-you-go credits. Anthropic’s Fable plan page now explains the different rules by plan.

24 July: Anthropic launched Claude Opus 5, included it in subscriptions and made it the Claude Code default. The email said it matched many Fable capabilities at half the price, while higher effort settings would consume more of the usage limit.

In less than three weeks, one model moved from temporary inclusion, to two extensions, to a planned paid-only future, to standard inclusion for some plans. At the same time, another model became the default and changed the usage calculation again.

A customer should not need a timeline to understand what their monthly subscription buys.

The problem isn’t paying for compute

Good AI costs money to run. I am not asking Anthropic to provide unlimited access to its most expensive models for a flat fee.

The trust problem comes from selling a subscription as productive capacity without giving customers a dependable floor for that capacity.

Anthropic’s Max plan costs US$100 or US$200 per month and is described as offering five or twenty times Pro usage. The same page says Anthropic may also apply weekly caps, monthly caps, or model and feature limits at its discretion.

When customers run out, usage credits continue at standard API rates, charged separately from the subscription. Customers can prepay for discounted bundles: pay US$45 for US$50, US$200 for US$250, or US$700 for US$1,000.

That can all be economically rational for Anthropic. It is still hard for a business to plan around.

A business cannot staff a process on “twenty times more” when the base amount is unclear, the model weights differ, promotions temporarily inflate the allowance and additional caps can appear at the vendor’s discretion. A consultant cannot confidently recommend a plan when this month’s answer may be wrong next month.

Promotions change customer behaviour

Temporary usage sounds generous. It also changes how people work.

Give a team 50% more Claude Code capacity for two months and they start more sessions, run more agents and hand over larger jobs. Give them a new model inside the same subscription and they learn where it fits. They build habits and processes around the capacity in front of them.

When the promotion ends, the business does not simply use less. It has to downgrade models, redesign the workflow, wait for a reset or turn on metered billing.

That is exactly what I have seen. Customers split automations into smaller jobs. They switch away from the strongest model to stretch the week. They stop asking what Claude can do and start asking how much of Claude they have left.

This is the wrong anxiety to breed in customers who are considering putting AI into day-to-day operations.

What businesses need from Anthropic

The answer is boring, which is usually what business infrastructure should be.

Publish a minimum included capacity for every plan in a unit customers can track. Give material changes a clear notice period. State whether a new model is a trial, a paid add-on or a permanent inclusion on launch day. Grandfather existing automations when the commercial treatment of an SDK or authentication method changes.

And make the usage dashboard good enough that a business can forecast next month from this month. We should be able to see consumption by model, product, user and workflow, with hard spending controls and an exportable history.

Anthropic can still test pricing. Every technology company does. But testing should produce clearer products over time. It should not make a paid plan feel provisional.

I use Claude every day. I recommend it to customers because the models are very good and the product is moving quickly. That recommendation now comes with more qualifications than it did a few months ago.

A brilliant model that I cannot budget or promise capacity for is a tool I keep at the edge of a business. A service with clear limits and stable terms is something I can build into the middle of one.

Anthropic says it wants Claude to do long-running work and operate across real business systems. To earn that role, it needs to give businesses something less exciting than another temporary usage boost.

It needs to give them certainty.