The Rationing: Anthropic's Second Confession

Six days after we published The Rationing, Anthropic posted a "Follow-up on usage limits" to r/ClaudeAI. They acknowledged users are hitting limits "way faster than expected." They adjusted peak-hour session windows. They said only about 7% of users are affected. The community response was immediate, voluminous, and almost uniformly negative. This is what the second confession looks like.

What Anthropic Said

The post, published April 2 on Reddit, is the most detailed public acknowledgment Anthropic has made about usage limits since the initial wave of complaints in late March. The key points:

Peak-hour throttling is now official. Anthropic confirmed they've adjusted 5-hour session limits specifically during peak hours, defined as 5am to 11am Pacific on weekdays. Outside those windows, limits are supposedly more generous. The framing is that this is a capacity management measure, not a permanent change — but no end date was given.

Demand drivers are named. Anthropic cited three factors: agentic AI features (Claude Code, tool use, multi-step tasks), the 1M context window (which consumes dramatically more compute per request), and general growth in the user base. This is the first time they've explicitly connected the 1M context window to the throttling. It's a significant admission — the flagship feature they marketed is part of why they can't serve it.

The March promotion is over. From roughly March 13 to March 28, Anthropic ran a promotion that doubled off-peak usage. That promotion ended. Users who calibrated their workflows to the promotional limits are now experiencing what feels like a 50% cut, because it is one.

The 7% number. Anthropic claimed only approximately 7% of users are affected by the tighter limits. This became the most contested claim in the thread.

What Users Said

The Reddit thread and surrounding posts paint a picture that is difficult to reconcile with "7% affected." Users across multiple tier levels — Pro, Max 5x, even Max 20x subscribers — reported the same pattern: session limits depleting far faster than the posted rates would suggest.

"I used up Max 5 in 1 hour of working."

This was one of dozens of similar reports. A 5-hour session limit exhausted in 60 minutes. Not through abuse or unusual usage patterns — through normal coding work with Claude Code. The math doesn't add up unless the actual per-message cost is significantly higher than what users were led to expect.

"Out of 30 days I get to use Claude 12."

This captures the cumulative effect. It's not just that individual sessions drain fast — it's that the cooldown periods between sessions make the tool functionally unavailable for the majority of a billing cycle. A user paying $100/month for Max who can productively use the tool 12 days out of 30 is paying roughly $8.33 per usable day. That's not a subscription. That's metered billing with extra steps.

"Typing 'Hello Claude' triggered a 4-hour cooldown."

Multiple users reported that trivially small interactions — greetings, single-sentence questions — were consuming disproportionate amounts of their session budget. This is where the story shifts from "demand exceeds supply" to "something is broken."

The Cache Bug

This is the thread that should worry Anthropic most. Users began investigating their own token consumption and found evidence of cache-related bugs inflating costs by 10 to 20 times the expected rate. The hypothesis: context caching — the mechanism that's supposed to reduce costs by reusing previous conversation context — was failing silently, causing every message to be processed as if the full context window were new tokens.

We covered the early reports of this in The Cache That Wasn't. What's new is the scale. More users are reproducing it. Some found that reverting to Claude Code version 2.1.34 — an older release — dramatically improved their consumption rates. Same prompts, same workflows, dramatically different token burn. That's not a demand problem. That's a software defect.

The community is now grappling with a question that Anthropic has not answered: how much of the "unprecedented demand" is real demand, and how much is a bug charging users 10x for the same work?

The Coverage Wave

This is no longer a Reddit drama. The Register, DevClass, PCWorld, and Axios have all published coverage of the usage limit controversy. The Register's piece was characteristically blunt. Axios focused on the business implications. The story has escaped the subreddit and entered the trade press — which means it has entered the narrative that enterprise procurement teams read.

Dario Amodei, Anthropic's CEO, made a public statement that there is "no hedge on earth" against overbuying compute. This is a remarkable admission from the CEO of a company valued at over $60 billion. Read it carefully: he's saying they can't buy enough GPUs to serve the demand they've created at the prices they're charging. The business model, at current pricing, does not close.

This is the same dynamic we described in The Rationing: subsidize adoption, create dependency, discover that the subsidy is unsustainable, then ration. Amodei is saying the quiet part — that the economics don't work — out loud.

What's Still Missing

Anthropic's follow-up post acknowledged the problem. It did not fix it. Here is what users still don't have:

Billing remediation. If a cache bug inflated token consumption by 10-20x, users were overcharged. Not in a "we feel overcharged" sense — in a "the meter was broken" sense. Anthropic has not announced any billing review, credit restoration, or retroactive adjustment for affected users. Every day without this is a day where users who were charged for a software defect remain uncompensated.

Specific fix dates. "We're working on it" is not a timeline. Users planning their work around Claude — freelancers with client deadlines, startups with runway pressure, teams mid-sprint — need to know when capacity will normalize. Anthropic offered no dates, no milestones, no commitments.

Transparent metering. Users still cannot see a real-time breakdown of their token consumption per message. They cannot audit their own usage. They cannot verify whether the cache is working. The system is a black box that tells you when you've hit the limit but won't show you why.

Acknowledgment of the bug. Anthropic's post attributed the limit pressure to demand growth and agentic features. It did not mention the cache bug reports that the community has been documenting for weeks. Silence on a known defect is not the same as denying it — but it's not reassuring either.

The Pattern

This is the fourth piece in our Rationing coverage, and the pattern is now unmistakable. The Rationing described the playbook: subsidize, addict, extract. Rationing Contracts showed that the legal framework permits all of this. The Cache That Wasn't found evidence that bugs were compounding the problem. The Vindication documented Anthropic's first round of acknowledgments.

Now we have the second confession. It's more detailed than the first. It names the causes more specifically. It still doesn't fix anything concrete. The confession-without-remedy cycle is itself becoming a pattern: acknowledge enough to defuse the news cycle, promise improvements without committing to dates, wait for the next wave.

The question for paying subscribers is simple: how many confessions before a fix?