When the AI Subscription Becomes a Bet: A Procurement Lesson

A controller at a German machine builder signs off on a 200-dollar monthly AI subscription. It is the tool the engineering team now uses to write quotation logic, and it costs less than a single contractor day. Five hours into a Monday coding session, the team hits a wall: roughly 15 percent of the weekly allowance is gone. That figure comes from the plaintiff in a lawsuit, not from an audited log. But the question it raises lands on the controller’s desk anyway. How do you budget for a cost line that nobody can convert into work units in advance?

That is the real story behind Kahn v. Anthropic (U.S. District Court, Northern District of California, filed mid-June 2026). The headlines read like AI gossip. For anyone who runs a business process on a large language model, it reads like a procurement lesson.

The four contracts nobody signed

Anthropic markets its Claude subscriptions as “5x” (100 dollars per month) and “20x” (200 dollars per month) more usage than the Pro tier. According to the Dataconomy report of 16 June 2026, the lawsuit alleges that the reference point behind both numbers was never disclosed. Five times what, measured how? The complaint, as summarised by Decrypt on 15 June 2026, traces the plaintiff’s own upgrade path: Pro in June 2025, Max 5x in January 2026, Max 20x in April 2026, each time chasing more headroom.

When you put a business process on a subscription like this, four things stay unwritten.

The first is how many credits you get per euro. A multiplier against an undisclosed base is not a quantity. The second is which models remain available. A subscription that quietly swaps the underlying model changes what your process actually buys. The third is how much of your allowance a single unit of work consumes. The fourth is whether model performance stays stable, the kind of dynamically throttled performance that users cannot see in any invoice line.

Exactly one number is contractual: the monthly bill. Everything else is trust.

A double cap that resets while you work

Engadget reported on 15 June 2026 that the complaint describes two limits stacked on top of each other: a five-hour rolling window and a separate weekly cap. The plaintiff’s filing quotes that “actual usage” fell “far below the advertised amount.” Two caps interacting, both resetting on different clocks, are hard enough for a power user to track. For a finance function trying to forecast a recurring cost, they are unforecastable by design.

There is a sharper detail. The weekly limits were not part of the launch offer. PYMNTS reported on 15 June 2026 that those weekly caps took effect on 28 August 2025, months after the April 2025 launch, and that Anthropic stated they would affect fewer than 5 percent of subscribers. The relevant point for a buyer is the timing. The cost variable your process depends on can be revised after you have committed to it, while the invoice stays flat and the capacity behind it shrinks.

This is not an Anthropic problem

It would be easy to read this as one vendor’s misstep. The wider market says otherwise.

GitHub announced in April 2025 that Copilot would move to usage-based billing, and the GitHub blog describes the shift from premium request units to token-metered AI credits priced at API rates. The change went live on 1 June 2026. Within days, The Register documented on 2 June 2026 developers reporting single changes that cost more than six dollars under metered billing, and threatening to leave. Same friction, different vendor: the price of a unit of AI work is moving from a flat promise to a variable nobody controls.

The regulatory frame does not close the gap. The EU AI Act, Regulation (EU) 2024/1689 of 13 June 2024, sets provider transparency obligations in Article 50. Those obligations cover disclosure about the AI system itself. They say nothing about commercial consumption transparency, about how many work units a euro buys. That gap is closed only by your own contract, which is the same logic that applies when a vendor switches off an interface after the fact.

What to check before you depend on an LLM

The lesson is operational, and it does not wait for the verdict. Three steps turn an AI bet back into a cost model.

Measure consumption per unit of work before going live. Run a one-week pilot with logging, take a representative task, and record how much of any allowance it actually consumes. If you cannot express the limit in work units, you do not have a budget line.

Treat the contract, not the marketing, as the reference. Use the API’s usage-precise token pricing as your yardstick. Sign a subscription only where the reference point is written down, not implied by a multiplier.

Define the exit path before you need it. Build a dual-vendor capability, meaning an abstraction layer behind which a second model provider is swappable, and document an API migration as the emergency exit if a subscription limit or a performance throttle makes the process uneconomic. The same principle decides who owns the data and the access when you want to leave. None of this is expensive, as long as you build it before the dependency hardens. This is precisely the work behind an AI rollout with a real cost model rather than a bet.

What the market figures don’t show

The serious counter-argument deserves a fair hearing. A subsidised flat-rate subscription is a genuine bargain. For 100 or 200 dollars you get compute that would run into four figures on the metered API. Anyone who wants a clean contract should simply take the API with its usage-precise token price. And a lawsuit that attacks the flat-rate model in court risks ruining the cheap plans for everyone else.

On price, that argument is correct. On planning, it misses the point.

A subsidised flat rate is cheap as long as the provider holds the limits stable, and that stability is exactly what is not guaranteed. Anthropic introduced the weekly limits on 28 August 2025, months after launch, and said fewer than 5 percent of subscribers would feel them, as Quartz reported on 15 June 2026. For a business, that means the cost variable a process was built on can shift without the invoice moving. The bill stays flat while the capacity behind it falls. Planning certainty is not the same thing as a low price.

The API is not the rebuttal to this article. It is the article’s answer. Metered consumption is the contract you can read; the subscription is the wager you cannot.

Frequently asked questions about LLM subscription risk

Does this mean I should stop using AI subscriptions?

No. It means you should calculate consumption per unit of work before a process depends on it. A subscription is cheap while the limit holds, but Anthropic introduced its weekly limits on 28 August 2025, months after launch. A figure that can be revised after the fact is a risk, not a planning variable.

Is the API really more predictable than a subscription?

On consumption, yes. Token billing is usage-precise and written into the contract, and GitHub Copilot moved to exactly this model on 1 June 2026. The total may run higher than a flat-rate plan, but you know in advance what each unit of work costs. That is the difference between a cost model and a bet.

Should I wait for the court to rule before acting?

No. Kahn v. Anthropic (N.D. Cal., filed mid-June 2026) is the plaintiff’s pleading, and the outcome is open. The operational consequence does not hang on the verdict. The due-diligence question, “do I know my reference point?”, applies to every vendor regardless of how this case ends.

What does an LLM exit path look like in practice?

Two components. First, an abstraction layer behind which a second model provider stays swappable, which gives you dual-vendor capability. Second, a documented API migration as an emergency exit, in case a subscription limit or a performance throttle makes the process uneconomic. Both cost little as long as you build them before the dependency takes hold.

Does this affect me if I only use AI for small tasks?

Then the risk is small for now, though the boundary moves quickly. As soon as a recurring business process sits on the tool, whether that is quoting, code, or analysis, the same question applies as with any supplier: what happens if the terms change and I cannot leave?

Next step

Can you express your AI tool’s limit in work units, or only in euros per month?

If your team has put a process on an LLM subscription and you want a second opinion on the dependency before it hardens, I am happy to compare your situation against the due-diligence checklist above. No pitch, no slide deck.

→ Book a no-cost intro call

→ Or read first: AI automation with a clear cost model · SAP API policy and integration risk

Sources and links: Dataconomy, 16 June 2026 · Decrypt, 15 June 2026 · Engadget, 15 June 2026 · PYMNTS, 15 June 2026 · GitHub Blog · The Register, 2 June 2026 · Quartz, 15 June 2026 · EU AI Act, Regulation (EU) 2024/1689, Article 50, 13 June 2024

Read more on pfisterer.xyz: Claude Opus performance degradation · SAP API policy 2026 · Data Act: access by design