Claude Max’s 20x promise draws a class action

An expanded class-action complaint filed September 8 says Anthropic sold its Claude Max plans on a “20x more usage” promise while not adequately disclosing a weekly cap. That cap sits on top of the five-hour sessions the multiplier is actually measured in, and developers paying $200 a month can still run into it.

The complaint was first reported by The Verge, and The New Stack has the details. It says Anthropic introduced weekly limits in late July 2025, months after Max launched in April, and kept marketing the $100 Max 5x and $200 Max 20x tiers on their 5x and 20x claims. The plaintiffs allege the constraints weren’t adequately disclosed when customers subscribed.

Per session, per week

Anthropic’s documentation says the Max multipliers apply “per session,” with usage limits resetting on a rolling five-hour window. Above that window sits a weekly usage cap across all models, and Anthropic says it may add other restrictions to manage capacity. Claude Code sessions count toward the same plan limits, so a day of interactive coding draws on the same weekly pool as everything else on the account.

Anthropic publishes no message count for the 20x tier. Tracked-usage estimates put Pro at roughly 45 requests per five-hour window and Max 5x at around 225. Going by the multiplier, Max 20x comes out on the order of 900, though model choice (Opus costs the most) and task complexity move that figure in both directions.

That variance is the core of the problem. A quick fix barely dents the allowance, while a more complicated job can burn through it much faster. Once the included usage is gone, a developer either waits for the reset or pays more to keep working. So the multiplier tells you how Max compares with Pro. It says nothing about how much coding $200 a month actually buys.

Anthropic’s pricing page still leads with “Choose 5x or 20x more usage than Pro*”, and the asterisk resolves to “Usage limits apply.” The same page lists a Fable model at “50% of weekly limits*” for both Max tiers, and adds that “Price and plans are subject to change at Anthropic’s discretion.”

A year of moving ceilings

The limits around that weekly cap have shifted several times since it arrived, and each shift changed what the multiplier was being measured against:

• April 2025: Max launches at $100 for 5x Pro usage and $200 for 20x. The top tier is pitched directly against OpenAI’s $200 ChatGPT Pro.
• Late July 2025: weekly limits arrive, per the complaint.
• May 6, 2026: Anthropic doubles the five-hour limits in Claude Code and drops peak-hour throttling.
• May 13, 2026: a promotional 50% increase to weekly limits begins. It is later extended more than once.
• June 14, 2026: Kahn v. Anthropic is filed in federal court in the Northern District of California. The plaintiff claims the $100 and $200 tiers deliver roughly 3.5x and 6-8x of Pro, against the advertised 5x and 20x.
• September 13-14: the temporary 50% bump is set to expire, and a permanent 25% increase over the standard weekly cap takes its place. Testing Catalog puts the net effect at a drop of roughly 17% against what subscribers had under the promotion.

Tuesday’s filing lands between the last two entries, less than a week before the weekly ceiling comes down from its promotional level.

Codex has the same problem

OpenAI faces the same variability with Codex. Its documentation tells subscribers that usage depends on the task, the model, and where the work runs. Codex can also draw from an allowance shared with other agentic products, and once that runs out, users can buy more credits. OpenAI has made Codex more autonomous too, so it can keep working while it waits for a developer to respond. Model choice changes the math as well: Astra costs 2.5x more per token than GPT-5.6, so the same allowance stretches much further on one model than on the other.

The mechanics aren’t identical to Claude Max, and the lawsuit doesn’t accuse OpenAI or any other provider of wrongdoing. Still, comparing AI coding subscriptions by monthly price tells a team little about how much work it gets done before it hits a limit.

OpenAI has also been testing outcome-based pricing with some enterprise customers. Under that model it charges only when an agent completes a task, instead of metering raw compute. That brings problems of its own: someone has to define “success,” and failed agent runs become the provider’s expense.

The label on the package

If the plaintiffs win, the case could set a precedent for how clearly AI providers have to explain usage restrictions before developers sign up, particularly when usage resets and what limits apply on top of the headline number. As coding agents take on larger jobs and run unattended for longer, the gap between what a plan advertises and what it delivers grows with the workload.

Anthropic has already given its answer to an earlier version of the case. In its motion to dismiss, the company argued that customers could reach information about the limits through hyperlinks during purchase. It compared those disclosures to a product label, which a buyer can read by turning the package over before deciding whether to buy.