Anthropic Sued Over Claude Max Usage Limits
Anthropic is sued over alleged Claude Max usage-limit shortfalls. What a $200/month subscription promise is worth — and how to hedge AI vendor lock-in.
On June 14, 2026, a proposed class action was filed in the US District Court for the Northern District of California: Kahn v. Anthropic PBC. The allegation, as CNET reported on September 3: subscribers to Claude's most expensive tiers — Max 5x at $100 per month and Max 20x at $200 per month — are getting far less usage than the plan names advertise. The complaint claims Max 20x delivers "just six to eight times the usage of Pro," and Max 5x delivers "just three-and-a-half times," instead of the promised multiples. Anthropic did not respond to CNET's request for comment. Nothing has been adjudicated; the suit was filed, not won.
That is the entire news peg. It is enough, because the claim is not exotic. "20x" was never a number. It was a shape — a multiple of an undefined base. Anthropic's own support pages describe usage limits as a "conversation budget" and reserve the right to cap usage "in other ways, such as weekly and monthly caps or model and feature usage, at [its] discretion." The lawsuit's core accusation is that the base is opaque by design: the complaint says Anthropic's site "is a black box, without any meaningful description of how usage is calculated." If that framing survives discovery, it will describe more than one AI vendor.
This post is not about the lawsuit. It is about what the lawsuit exposes: a vendor-risk pattern every founder running agents on subscription tiers should assume applies to their stack, and the contract and routing posture that hedges it.
The vendor-promise pattern
The pattern in the complaint will be familiar to anyone who has run production workloads on a premium AI subscription:
- The plan is sold as a multiplier. Max 20x costs $200/month and promises twenty times the usage cap of Pro ($17–$20/month). Multiples are intuitive and unfalsifiable at purchase time, because the base they multiply is itself undefined.
- The limit is not a number. Anthropic's documentation describes per-session and weekly limits but, per the complaint, no post states how many prompts each plan allows in the relevant windows. The plaintiff reported burning nearly 20% of his weekly allocation in a single five-hour coding sprint — on a plan marketed as twenty times the base.
- The resolution is spend. Subscribers who hit a rate limit are prompted to buy more credits. Anthropic's Max plan terms reserve discretionary caps. The complaint alleges a "bait and switch" with supporting Reddit testimony, and seeks class status for Max purchasers since April 2025, with the amount in controversy stated as exceeding $5 million.
Whether or not the court agrees, the structural observation stands: when the unit you are buying is undefined, the vendor holds every adjustment lever, and the upgrade path from an unexpected cap is always more spend. We already watched one version of this play out with Anthropic: the company closed the third-party proxy loophole that routed heavy Claude Max usage through its own Agent SDK, which broke workflows built on the assumption that a subscription multiplier meant an API multiplier. That migration was covered here when it happened — and it was a quieter event than a lawsuit, with the same lesson.
What $200 per month does and does not buy
Strip the marketing and a premium AI subscription buys three concrete things: a seat, a soft usage envelope, and a billing relationship. What it does not buy is a capacity commitment. There is no SLA on "20x," no definition of a session, no published prompt budget. The complaint's claim that Max 20x behaves like 6–8x of Pro is, if true, a disclosure problem. But even if Anthropic's numbers were exactly right, the founder's problem is unchanged: a plan denominated in an undefined unit is a plan the vendor can resize at will, and the only recourse inside a single-vendor stack is to pay more.
This is the quiet version of AI vendor lock-in, and it is more common than the dramatic one. The dramatic version is a vendor shutting off access or tripling prices. The quiet version is a limit that moves under your workload — a weekly cap that tightens as your agents' usage grows, a "fair use" policy interpreted against your heaviest workflow, a multiplier that meant one thing when you sized your stack and another after a pricing-page edit. You find out when the prompt to buy credits appears, not before.
For an agent stack, the stakes are higher than for a seat-and-keyboard user. An individual coder who hits a cap loses an afternoon. A production agent hitting an unexpected cap mid-run loses the run — or, worse, silently degrades if the failure mode is a fallback nobody tested. Usage limits are an infrastructure dependency, and infrastructure dependencies deserve hedging, not trust.
The hedge: contract, measure, route
Three controls move this from trust to engineering. None requires predicting any vendor's behavior.
Contract for the unit, not the multiple. If a vendor's paid tier matters to your operations, the negotiation is not over price — it is over the definition of the unit you are buying: what a session is, how usage is calculated, what the weekly and monthly caps are, and what happens to your workloads when a cap is reached. A multiple of an undefined base is not a term; it is an impression. Some vendors will refuse to define it. That refusal is itself information, and it should feed directly into the next two controls.
Measure your actual multiplier. The plaintiff's case rests on a comparison any subscriber can run: track consumption across tiers and time, and compute the effective multiple yourself. For an agent stack this is straightforward — log tokens and requests per workflow, per model, per day, and compare against the plan's marketing. If your "20x" plan sustains 8x of Pro's throughput on your workload, you now know the real price of your capacity, and you can price the alternative accurately. Tacavar's own tiered gateway exists partly because of this: route by task, fall back by cost, measure everything. The full cost anatomy of a multi-model stack is covered in the multi-model founder stack piece; the short version is that first-party routing numbers make vendor promises checkable, and checkable promises are the only kind worth building on.
Route so no single limit is fatal. The terminal risk of single-vendor dependence is not that the vendor behaves badly — it is that your operation has no answer when it does. A multi-model routing posture converts one vendor's limit from an outage into a re-route. That does not mean running every model for every task; it means your stack can degrade across at least two providers per critical workflow, with the fallback path exercised in production, not discovered in an incident. If you want the canonical recent example of what happens when a vendor silently changes the rules of a subscription, the Claude Max proxy-loophole closure is the case study: workflows built on an undocumented assumption stopped working the day the assumption did.
Frequently Asked Questions
What is the Claude Max lawsuit about? Kahn v. Anthropic PBC, filed June 14, 2026 in the Northern District of California, alleges false advertising: that Claude Max 5x and Max 20x deliver roughly three-and-a-half and six-to-eight times the usage of Pro, respectively, instead of the advertised five and twenty times. The complaint seeks class action status for Max subscribers since April 2025. As of this writing it is an allegation, not a ruling.
What are Claude Max usage limits? Per Anthropic's documentation, usage limits are a "conversation budget" governing how much a subscriber can use Claude within set windows, including per-session and weekly limits. The lawsuit's central claim is that Anthropic does not define the underlying unit — how usage is calculated or how many prompts a plan allows — leaving the advertised multiples unverifiable.
What is AI vendor lock-in? Dependence on a single AI provider for production workloads, such that a pricing change, usage-limit change, or access change becomes an operational crisis rather than a procurement decision. The quiet form — undefined units and movable limits — is more common than outright access loss, and it is hedged with defined contracts, first-party usage measurement, and multi-model routing fallbacks.
Is a multi-model stack more expensive than one premium subscription? Usually in raw spend, often not in cost per successful task. Subscription seats are cheap until a cap bites; API and routed usage is metered but predictable. First-party routing data lets a founder compare the real cost of capacity across both — which is the point.
You built it. We optimize it.