Anthropic expects to earn $18 billion this year. Some analysts estimate they have already exceeded their target ahead of schedule. But clearly, they are not satisfied.

Early this morning, a developer named George Pu posted his discovery on X: “Anthropic (Claude’s parent company) just removed Claude Code from the Pro plan. Pro users who want to use this feature now have to upgrade to the Max plan, with a minimum monthly fee of $100.”

This means that to maintain the same permissions, the price has increased fivefold. Since there are no blogs, no emails, no changelogs, Anthropic quietly updated the webpage in a subtle way, attempting to raise prices.

This tweet quickly spread. Screenshots of the pricing page began circulating on Reddit, Hacker News, and developer groups. The community repeatedly confirmed on their screens: under the Pro column, the line for Claude Code now shows a cross, with only Max 5x ($100/month) and Max 20x ($200/month) still checked. Previously, Anthropic’s support documentation stated “Using Claude Code with your Pro or Max plan,” but now it reads “Using Claude Code with your Max plan.”

The community naturally expressed strong dissatisfaction. As public pressure built, Anthropic’s growth lead Amol Avasare had to respond on X. He claimed: “This is just a small-scale test targeting about 2% of newly registered Pro users; existing Pro and Max subscribers are unaffected.”

Later, he posted a longer explanation in the same thread, roughly saying: When Max was launched a year ago, it didn’t include Claude Code, and Cowork didn’t exist yet. Long-running agents weren’t part of the daily workflow. The original design goal for Max was heavy chat users, nothing more.

But after Opus 4 was released, the usage of Claude Code exploded. Asynchronous agents became a daily tool for developers. “Per-user usage increased significantly, and our current pricing scheme wasn’t designed for this kind of usage.”

He finally assured: if the final plan affects existing subscribers, they will be notified well in advance, “the message will come from us, not screenshots on X or Reddit.”

Months of continuous tightening of compute power

Avasare’s explanation sounds reasonable, but taken together, it points to a more fundamental issue: Anthropic may be under serious compute pressure.

Every coding session with Claude Code consumes far more tokens than a normal conversation, sometimes by an order of magnitude. After Opus 4 was released, user demand surged. Agents started running hours-long asynchronous tasks. As Cowork brought more people into intensive use cases, the marginal cost of the $20/month plan began to spiral out of control.

From a business perspective, this isn’t just a minor annoyance of “users using more,” but a fundamental mismatch in the pricing model after rapid product capability evolution. Max was originally designed for heavy chat, but now it needs to support continuous-running programming agents. The compute costs for these two scenarios are in completely different leagues.

Avasare used a phrase: “engagement per subscriber is way up.” In business terms, it means: for the same money, users are now consuming many more resources.

This isn’t the first time. Looking back over the past few months, Anthropic’s “compute tightening” actions are quite frequent.

In August 2025, Anthropic announced weekly usage caps for Pro and Max users, effective from August 28. The official reason was that some users violated terms by sharing and reselling accounts to bypass limits. The statement said, “It is expected to affect less than 5% of subscribers.” Weekly caps marked the first time high-paying users faced strict weekly limits.

Later, from late August to early September, things got more complicated. Many developers reported on Reddit and X that Claude’s code generation quality plummeted—models started ignoring their plans, producing gibberish randomly, and Claude Code’s task execution became unreliable. Anthropic eventually admitted this was a technical bug caused by an upgrade to the reasoning stack, which led to a continuous degradation of Claude Opus 4.1 over about 56.5 hours, and they published a detailed post-mortem.

This bug incident quickly overlapped with a larger controversy.

In early March 2026, a senior AI director at AMD, Stella Laurenzo, posted a detailed analysis on GitHub. Based on 6,852 Claude Code session files, 17,871 thinking blocks, and over 230k tool calls, she concluded:

Claude Code has become less intelligent. This analysis resonated strongly within the developer community.

Boris Cherny, head of Claude Code, responded afterward, denying the core conclusion that “the model has worsened.” He explained that Opus 4.6 switched to an “adaptive thinking” mode on February 9, and on March 3, the default effort level was set to medium (level 85). The official stance was that this was “the best balance between intelligence, latency, and cost.” But he also admitted that about 7% of users would hit session limits they previously wouldn’t, and this adjustment mainly targeted peak-time Pro users, with Team and Enterprise unaffected.

From March 13 to 27, 2026, Anthropic ran a promotion called “doubling usage outside peak hours”: from 2 p.m. to 8 a.m. the next day (Eastern Time), and all weekend, usage limits automatically doubled for all plans—Free, Pro, Max, Team.

On the surface, it looked like a benefit event. But in context, the reason for doubling during off-peak hours was because the compute capacity was under strain during peak times. Doubling the quota during idle periods didn’t affect congestion, while also creating incentives for users to upgrade to higher plans.

Then, on April 4, Anthropic announced that third-party agent frameworks could no longer operate within the usage limits of Pro and Max subscriptions. If users wanted to continue calling Claude via external frameworks like OpenClaw, they would need to pay extra per call.

And then there was the disappearance of Claude Code from the Pro pricing page.

A company like A is definitely not doing this for the last time

Returning to Avasare’s tweet. He said, “The removal of Claude Code from the Pro pricing page is just a small-scale test targeting about 2% of newly registered Pro users,” but this explanation triggered widespread skepticism in the community.

The most direct question came from developer Simon Willison: “Everyone I know has seen the new pricing page, and the Internet Archive has already saved a screenshot.” If it only affects 2% of users, why was the entire public pricing grid changed? Why was the support documentation updated across the site?

More intriguingly, when a reporter asked Avasare why the pricing page and support docs were changed so comprehensively, and why a test affecting only 2% of users couldn’t explain this scope, he didn’t respond. Anthropic’s spokesperson also did not answer this follow-up.

Another inconsistency noted was that Claude Cowork was not removed from the Pro plan, but from a product perspective, Cowork is essentially just Claude Code under a different brand name. As one developer put it, “Cowork is Claude Code in a less intimidating disguise.”

The development unfolded as: hours after the controversy erupted, Anthropic quietly restored the website and support documents to their original state. No explanation, no announcement. Just like the change itself, it was done silently.

Looking at all these actions together, it’s clear that Anthropic’s current approach is a calculated test:

They are subtly pushing the boundaries of compute usage for paying subscribers. Doubling during weekends is meant to steer demand toward off-peak times; lowering the default effort level is to save compute during busy hours; banning third-party agent frameworks is to close loopholes for quota arbitrage; and removing Claude Code from the Pro page is testing user reactions to more aggressive re-pricing schemes.

Avasare himself admitted: “Our current scheme wasn’t designed for this kind of usage.” That’s a rare frank admission. It indicates that Anthropic has, to some extent, realized that their current pricing structure can’t sustain the compute demands brought by Claude Code.

What will happen next, we don’t know.

But perhaps, as Ed Zitron wrote: “I don’t think this will be the last time Anthropic does something like this.”