<h2>What’s happening</h2>
<p>Claude Code users on paid plans — including the $200/month Max tier — are hitting their usage limits dramatically faster than expected. Some report exhausting a 5-hour session quota in under 20 minutes. Others say a single prompt is consuming 10% of their session allowance when it should cost closer to 0.5–1%.</p>
<p>Anthropic has acknowledged the problem publicly, calling it a top priority, though as of writing they haven’t identified a definitive root cause.</p>
<h2>Why it’s happening (probably)</h2>
<p>Several things appear to have converged at once:</p>
<p><strong>Peak-hour throttling.</strong> Anthropic quietly reduced usage limits during weekday peak hours (5am–11am Pacific, 1pm–7pm GMT) to manage GPU capacity. The change was announced via a single employee’s post on X rather than any official channel, and affected roughly 7% of users who hit limits they’d never hit before.</p>
<p><strong>A promotion ended.</strong> A period of doubled limits expired, so users who’d grown accustomed to that headroom suddenly found themselves working with half as much.</p>
<p><strong>Possible cache bugs.</strong> A community member claims to have reverse-engineered the Claude Code binary and found two bugs that break prompt caching, silently inflating token costs by 10–20x. This would explain why taking a break and resuming work can cost more than working continuously — the cache expires and everything gets reprocessed from scratch.</p>
<p>The combination is brutal. Automated workflows are especially vulnerable: a rate-limit error mid-run can silently drain an entire daily budget in minutes before anyone notices.</p>
<h2>The opacity problem</h2>
<p>Anthropic doesn’t publish exact token allocations for its plans. Instead, they use relative multipliers — the Max plan is described as “5x” the Pro plan — which tells you almost nothing useful when you’re trying to plan a workday around your remaining quota.</p>
<p>This makes it genuinely hard to budget work. You can’t tell how much headroom you have, you can’t predict when you’ll hit a wall, and when something goes wrong (like a cache bug inflating costs 10x), you have no way to detect it until you’re already locked out.</p>
<p>Some users have resorted to downgrading to older versions of Claude Code as a workaround. Others are looking at OpenAI Codex.</p>
<h2>The bigger picture</h2>
<p>Anthropic is caught between competing demands on its GPU infrastructure: researchers training new models, product teams building features, and a rapidly growing enterprise customer base. The $200/month Max plan is reportedly subsidized by something like 25x, so the economics of unlimited usage were never going to hold forever.</p>
<p>The rate limit changes themselves are arguably defensible. The communication around them isn’t. There was no advance notice, no official announcement, no dashboard showing current limits or usage. Developers found out when their tools stopped working.</p>
<p>The contrast with how OpenAI handles similar situations — frequent limit resets, public communication, some acknowledgment that developers are trying to build things — is hard to ignore. Anthropic has a reputation for building impressive models while being genuinely bad at talking to the people who use them.</p>
<p>For now, if you’re hitting limits faster than expected, the practical workarounds are: work outside peak hours, avoid long breaks mid-session (to preserve cache), and watch for the bug fix Anthropic says is coming.</p>


Collections

Claude Code users on paid plans, including the $200/month Max tier, are exhausting usage quotas far faster than expected — some burning through a 5-hour session in under 20 minutes. Three factors appear to be converging: Anthropic quietly reduced limits during peak hours with minimal communication, a doubled-limits promotion expired, and a suspected cache bug in the Claude Code binary may be inflating token costs by 10–20x. Anthropic has acknowledged the issue but hasn't identified a root cause. The core frustration is opacity: Anthropic doesn't publish token allocations, uses vague relative multipliers, and provides no usage dashboard. Workarounds include working outside peak hours, avoiding mid-session breaks to preserve cache, and downgrading to older Claude Code versions. The incident highlights a recurring criticism that Anthropic builds strong models but communicates poorly with developers.

Anthropic throttles Claude subscription limits during peak hours