Claude's rough Tuesday: multiple models degraded in a live incident while the Claude Code usage boost expires tomorrow night

Claude's rough Tuesday: five models degraded, a fix in two hours, and the 50% Claude Code boost dies tomorrow night

Claude developers spent Tuesday afternoon watching two clocks run down at once. The first started at 16:20 UTC, when Anthropic's status page opened an incident for "degraded performance for multiple models": elevated errors on Claude Opus 5, Sonnet 5, Fable 5, Mythos 5 and Haiku 4.5, across claude.ai, the Claude API, Claude Code and Claude Cowork 1. The second runs out at 11:59 PM PT on August 19, when the promotion that has held Claude Code weekly limits 50% above standard since May 13 ends, and the limits, in Anthropic's words, "return to their standard levels" 2. That expiry converts to 06:59 UTC on August 20, which leaves 38 hours and 39 minutes between the incident's opening post and the end of the boosted era (our arithmetic).

Thread A: the outage that was, five models to a fix in two hours

The incident's own timeline, as posted:

• 16:20 UTC: the incident opens, and an update the same minute cites elevated errors on Claude Mythos 5, Claude Fable 5, Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5, and other Claude models 1.
• 17:12 UTC: the investigation narrows to "elevated errors on requests to Claude Opus 5" 1.
• 18:26 UTC: "A fix has been implemented and we are monitoring the results" 1.

Two hours and six minutes separated the opening post from the fix, which landed with 36 hours 33 minutes left on the promotion clock (our arithmetic).

Thread B: tomorrow night the Claude Code ceiling drops by a third

The promotion covered Pro, Max and Team plans plus legacy seat-based Enterprise seats, and it touched only Claude Code's weekly limits, across the CLI, IDE extensions, desktop and web. Five-hour limits were never boosted; Free plans and consumption-based Enterprise seats were never included 2. Anthropic notes "We've extended this promotion," and a launch-day Hacker News thread shows why that sentence exists: it put the original end date at July 13 3. Launch to expiry is exactly 14 weeks, 98 days: 61 as originally scoped, 37 added by the extension (our count). When it lapses, the standard limit is the promotional ceiling divided by 1.5, two-thirds of it, a one-third cut from what developers have planned around since mid-May. That figure is our arithmetic, not Anthropic's; the help page says only that limits return to standard.

The squeeze: uptime and quota devalued inside one window

Our read: a coding subscription is paid for in two currencies, how often the models answer and how much of them you can spend. Tuesday devalued both. The degradation spanned claude.ai, the Claude API, Claude Code and Claude Cowork, and Claude Code is the one surface whose limits shrink tomorrow night 1. Hacker News processed the two events as one story, in parallel threads posted 34 minutes apart: the outage thread at 16:28 UTC and the front-page limits thread at 17:02 UTC 45. Treat the thread pairing as sentiment, not corroboration of impact. The exposure is real regardless: workflows tuned to the 1.5x ceiling spent their final boosted hours being stress-tested during a live degradation.

What to audit before 11:59 PM PT tomorrow

• Run /usage in the Claude Code CLI, the check Anthropic's help page itself points to, and compare your weekly burn against two-thirds of the ceiling you have been planning around 2.
• List the scripts that pin Claude Opus 5, the one model the 17:12 UTC update named 1.
• Watch for whether Anthropic ever acknowledges the calendar it built.

The next day and a half answers two questions: whether the fix holds, and whether standard limits still fit the workflows developers built all summer on the boosted ones.