SemiAnalysis's hands-on testing claims Anthropic is running an A/B gray test on Claude subscription quotas

Recently, quite a few users have complained that Claude's quota is suddenly being consumed at a highly abnormal rate.

Just as everyone was wondering whether they were experiencing a collective hallucination, the well-known semiconductor and AI research firm SemiAnalysis dropped hard evidence.

They posted on X to sound the alarm, confirming that users' perception was not wrong at all: Anthropic really is quietly running an A/B gray test on Claude subscription quotas in the background.

On the same day, SemiAnalysis ran an Opus 5.5 agent workload test on two completely identical $200 Max 20x accounts, and the results revealed a very clear difference.

The test group's account B was given a 5-hour window quota of 2.076 billion tokens, 19% higher than the normal account A's 1.74 billion tokens.

But looking at the more critical weekly total ammunition, account B shrank from the normal 6.579 billion tokens to 5.266 billion tokens, wiping out a full 20%.

Converted into monthly API-equivalent value, the normal account A can generate $11,726 worth of compute value per month, equivalent to 59 times the $200 monthly fee.

Meanwhile, the unlucky account B pushed into the gray-test group saw its monthly equivalent value fall to $9,385, with the subsidy multiplier dropping to 47 times the monthly fee—heavy users are effectively having a full 20% of their quota taken away each month.

No wonder that over the past couple of days, countless Vibe Coding folks who pay for the $100 and $200 plans suddenly found that the Opus 5.5 in their hands had completely changed flavor.

People used to joke that ChatGPT's catchphrase was "You're absolutely right," but with Claude Code today, its catchphrase has become "You've reached your usage limit."

Another X user reported that last week they could barely feel any limits, but today, before they had even run many heavy tasks, 23% of their quota was already gone.

Another user named Priya was even more helpless: their daily work had just begun, and the 90% progress bar was already at the bottom.

What's even more interesting is the official damage-control posture right away.

Facing questions from across the internet, Claude Code staff member Lydia Hallie came out to reassure everyone, explicitly stating that there had been no adjustment to quotas and suggesting that users who felt consumption was fast should run the feedback command and DM her their ID for investigation.

Right after that, she offered a set of explanations shifting the blame, asking users whether they often left in the middle of long conversations for more than an hour, causing the context cache to expire, or whether they had opened too many sub-agents in a session.

Claude Code's core developer Boris Cherny has also just stated that the team is urgently investigating the cause and will give everyone an explanation soon.

The most unbearable part of the whole affair is that just half a month ago, SemiAnalysis had given Anthropic a highly valuable good-guy card.

English text is being prepared.

After a round of extreme stress testing, among the $200-tier plans, the API-equivalent value provided by the Claude subscription using Opus 5.5 was about 5 times that of OpenAI's same-tier plan using GPT-6.1 Sol.

The direct trigger for the gap widening to a runaway 5x lead was that neighboring OpenAI acted earlier and with even less restraint.

OpenAI had previously halved the token quotas for the various models in its $200 Pro plan outright, and even erased the relative-multiplier promotional language from its official pricing page.

After cutting the $200 plan, OpenAI turned around and launched a new $500-per-month tier.

But actual testing showed that for the extra $300, the top-tier Astra model's quota was only about 21% higher than the pre-adjustment $200 plan, and the real gimmick became the ultra-fast output mode of 300 tokens per second.

In short, the hundreds of extra dollars users pay do not buy much additional quota; a large part of it is simply paying for output speed. Now, the once-straightforward Anthropic has ultimately taken the same approach.

The reason may be that, in the eyes of vendors such as OpenAI and Anthropic, the heavily subsidized monthly subscription model has become the biggest drag on profit margins.

According to SemiAnalysis's estimates, the subscription business contributes only about 10% of Anthropic's total revenue but consumes more than 40% of its inference compute.

But does the large-model business really have to be sustained by continuously shrinking user quotas to maintain gross margins? Turning our gaze back home, looking at the engineering answers just delivered by domestic model vendors presents a completely different picture.

Just in the past couple of days, DeepSeek published the technical practice of large-scale inference deployment for V4.1 Flash, and officially open-sourced its cloud service inference framework DeepInfer. Using common 8-GPU machines for prefill and 64- to 128-GPU supernodes for decode, DeepSeek achieved a daily throughput of 155 trillion tokens for a single cluster.

There are two very hardcore data points in this report.

The first is that, at the system design level, DeepSeek supports retaining KVCache context caches for more than 72 hours, solving the cost challenge of long-cycle context reuse with solid engineering optimization.

The second data point directly lays bare the money-making capability of this business: even calculated at off-peak low prices, the overall revenue of DeepSeek's inference system reached $38 per GPU per hour.

Doing the simplest math, a single GPU can generate more than $900 in revenue in 24 hours a day, fully proving that as long as the underlying inference architecture and scheduling optimization are done well, the AI model business can absolutely achieve extremely high gross margins.

In the mobile internet era, the marginal cost of software and content services was almost zero; after users signed up for a monthly membership, no matter how frequently they used it, the extra bandwidth cost the platform paid was negligible.

And what makes users most uncomfortable may not be the price increase itself, but the lack of basic decency and transparency in the whole process.

Vendors do not disclose specific consumption standards, only giving a vague percentage progress bar, either secretly degrading model intelligence or quietly running A/B tests in the background to lower quotas.

Users pay real money for their subscription fees on time every month, yet the service standard they buy may be tampered with at any time without their knowledge. This kind of trust overdraft, treating users as suckers, is probably harder to make up for than the consumption of compute.

This article comes from the WeChat public account "APPSO", author: the one who discovers tomorrow's products, published by 36Kr with authorization.