A Claude Max user, who pays $100 a month, generated $5,600 in actual API costs in a single billing cycle. It is just an (extreme) example of the enormous gap between the price of such a plan and its spending potential. Other analyzes place it on the range of between 1,000 and 5,000 dollars.
Anthropic has just cut off access to third-party tools (such as OpenClaw) to your subscription plans. In addition to for protecting your fenced gardenis also the inevitable arithmetic consequence of selling tokens at open buffet prices when real consumption has skyrocketed between 10 and 100 times.
Why is it important. The business model that has financed the explosion of AI for programmers (flat rates, unlimited access, subscription model) has been built on a statistical fiction that development agents have destroyed.
What is assumed when a flat rate is sold in any area is that light users subsidize the most intensive ones and the averages are sustained. That works in the data rates of a telecom, in the gyms, in Netflix and in an all-you-can-eat buffet. But it doesn’t work when any user can become, from one day to the next, a beastly consumer of computing. And that’s exactly what happens when an agent comes into play. And since December we are in the era of the agents.
Between the lines. The person who has best explained this problem is Luo Fuliresponsible for the team MiMo model on Xiaomi and ex-DeepSeek, who has published a long tweet which has moved especially in Chinese technological circles.
Your diagnosis is that third party tools like OpenClaw are not optimized to reuse Claude’s context cache, so each query regenerates context windows of over 100,000 tokens from scratch. The actual number of requests per query is several times higher than what Claude Code himself would generate. “That’s not a gap. It’s a crater,” Luo said.
The backdrop. In China, AI programming plans have become one of the most sought-after technology products on the market. Alibaba Cloud quotas They usually sell out in the first hours of the work dayand Tencent’s appear as “unavailable” permanently.
Developers go so far as to set morning alarms and write scripts self-purchase to get monthly access. The demand is real, but the underlying economics point in the opposite direction: subscription plans, designed for a world where each interaction consumed a few hundred tokensnow absorb agent workloads that consume 10 to 100 times more per task.
Yes, but. Anthropic has not closed access to third-party agents. You have moved them to another invoice: from flat rate to pay-as-you-go API. The measure includes a one-time credit equivalent to the monthly price of the plan and discounts of up to 30% for those who pre-purchase “extra use” packages.
The problem is that for many independent developers, the jump in costs, potentially tenfold, makes the use of agents no longer economically viable. Some have already announced that they will migrate to other models. Someone has to pay for that party, and until a new, more efficient solution arrives, no one knows who it will be.
The big question. If flat rate can’t survive actual agent usage, what pricing model can? Luo Fuli believes that economic pressure will eventually force third-party tool developers to optimize their context management and maximize cache reuse.
He may be right and his approach is logical. But in the meantime, the entire industry is operating with a business model whose math doesn’t add up, and the question of who will absorb the difference (suppliers, developers or end users) remains unanswered.
In Xataka | Companies have been investing in AI for years. The problem is that many projects are not producing results.
Featured image | Fotis Fotopoulos
GIPHY App Key not set. Please check settings