Anthropic limits Claude 5-hour sessions during 5am-11am PT peak window

Anthropic's announcement is narrow but consequential for heavy users: the company is changing how fast users advance through the 5-hour session bucket during weekday peak demand, specifically 5am-11am PT / 1pm-7pm GMT. The weekly cap does not change, but the session meter now depletes faster in that window for free, Pro, and Max subscribers.

The practical implication is that session-based workloads and weekly quotas are now decoupled more sharply. Screenshots circulating in the discussion show users reaching 100% of the current session while weekly usage was still only partly consumed, which matches the quota screenshots discussed in the broader thread.

Anthropic also gave one technical explanation for why some cases may look worse than expected. In the follow-up reply, a company employee said users are often hitting "expensive prompt cache misses" when they resume long conversations on "million context," and said a separate thread on avoiding that is coming.

User reports describe two overlapping problems: ordinary demand shaping from the new peak-hours rule and intermittent Claude Code reliability issues. In the most-cited example, one developer said Claude Code was "nearly unusable yesterday," with "major outages all day," and reported hitting the $200 Max-plan limit in less than an hour in that incident report.

A later update from the same user showed the opposite pattern once service normalized: "6 Claude Code agents" running Opus 4.6 in parallel used only 11% of the 5-hour session after 30 minutes, according to the recovery post. That contrast suggests at least some of the worst burn rates were amplified by service instability rather than steady-state usage alone.

Another user pushed back on Anthropic's cache-miss explanation with a more controlled anecdote. In the reply, they said they were using fresh sessions under 200k context on Pro, watched the 5-hour meter climb "5%" on each refresh, hit the limit in under 30 minutes during the morning window, then saw off-peak usage run for more than two hours while staying below 30% of the same bucket.

Some developers are already routing around the issue. One said they were "switching to Codex with GPT 5.4 for the day" in a migration post, and another later said Claude Code outages had kept pushing them into Codex, where GPT-5.4 xHigh "keeps delivering," per the comparison post. The engineering takeaway is simple: Claude's session economics now vary materially by time of day, and recent Claude Code instability made that change much more visible to users running long, token-heavy coding sessions.