Live archive · independent footnotes

Claude ships.We keep the footnotes.

ClaudeWatch is an independent, mobile-first field guide to Claude mishaps: quality dips, capacity plots, harness bugs, and classic slop patterns. Sourced in public from @claudeai and @ClaudeDevs, plus community-shaped cases. Satire included. Affiliation not included.

Read the timelineSend a tip
38filed cases
2official X sources watched
95desktop teal forever

Chronology

Issue timeline

Official mishaps first, oldest to newest when you open the full page. Home shows the latest pulse.

Full timeline
2026-08-19

Opus 5 and Haiku 4.5, identified, fixed, next slide

Cause identified. Cause unnamed. Users already know the error code.

Status page identified a cause of elevated errors on Claude Opus 5 and Claude Haiku 4.5, implemented a fix, and marked the incident resolved the same morning UTC. Affected claude.ai, API, Claude Code, and Claude Cowork. No public root-cause detail beyond that sequence.

outagereliabilitystatus.claude.com · 19 Aug 2026
2026-08-18

Weekly Code limits stay 50% higher, now through a newer month

Permanent is the destination. August 31 is the layover they keep booking.

@ClaudeDevs extended the 50% increase to weekly Claude Code limits through August 31 for the same plan families previously covered. The post hopes to make the change permanent, and notes capacity may stay tight from strong demand.

capacityproduct-trust@ClaudeDevs · 18 Aug 2026
2026-08-18

The family portrait of elevated errors, then just Opus

Start with the whole catalog. End with one model and a tidy UTC range.

Status page investigated degraded performance across multiple models, listing Mythos 5, Fable 5, Opus 5, Sonnet 5, Haiku 4.5, and others. Later updates narrowed to Opus 5. Official impact window 16:11 to 18:23 UTC. Resolved same day.

outagereliabilitystatus.claude.com · 18 Aug 2026
2026-08-17

Opus recovered. Sonnet stayed for the second act.

Two models, one incident, staggered dignity.

Status page investigated elevated errors on Claude Opus 5, then reported Opus operational while Claude Sonnet 5 stayed degraded. Fix implemented and incident resolved the same afternoon UTC. Affected claude.ai, API, Claude Code, and Claude Cowork.

outagereliabilitystatus.claude.com · 17 Aug 2026
2026-08-16

Login first, then the rest of the product remembered it existed

If you cannot sign in, every other feature is a rumor.

Status page investigated an authentication issue for claude.ai, Claude Code, and Claude Cowork, then widened to degraded performance on those surfaces plus platform.claude.com and the Claude API. Fix deployed and marked resolved the same evening UTC.

outagereliabilitystatus.claude.com · 16 Aug 2026
2026-08-07

The safety classifier finally stopped billing you for the lecture

A small number of extra tokens, until someone notices the meter.

@ClaudeDevs said Auto mode runs a classifier on every tool call and that as of 7 Aug 2026 that classifier overhead no longer counts toward usage limits on Pro, Max, and Team. Enterprise and API get the change once Auto mode is default for those users.

billingproduct-trustagent-harness@ClaudeDevs · 7 Aug 2026

Field guide

Slop patterns

Names for the recurring bits so you can point without writing a novel.

Taxonomy

Sycophancy

Agree first. Verify never. Complimentary empty calories.

Empty confidence

Tone of a senior. Footnotes of a rumor.

Code that lies

Compiles, demos, then invents the rest of reality.

Capacity theater

Demand was unpredictable, said every packed launch.

Harness mishaps

Model fine. The robot body typing into git had other plans.

Template prose

Six headings, zero decisions, infinite politeness.

Also filed

Latest cases

All cases
2026-08-19infrasev 4/5official

Opus 5 and Haiku 4.5, identified, fixed, next slide

Cause identified. Cause unnamed. Users already know the error code.

Status page identified a cause of elevated errors on Claude Opus 5 and Claude Haiku 4.5, implemented a fix, and marked the incident resolved the same morning UTC. Affected claude.ai, API, Claude Code, and Claude Cowork. No public root-cause detail beyond that sequence.

outagereliabilityclaude
2026-08-18X @ClaudeDevssev 3/5official

Weekly Code limits stay 50% higher, now through a newer month

Permanent is the destination. August 31 is the layover they keep booking.

@ClaudeDevs extended the 50% increase to weekly Claude Code limits through August 31 for the same plan families previously covered. The post hopes to make the change permanent, and notes capacity may stay tight from strong demand.

capacityproduct-trustclaude
2026-08-18infrasev 4/5official

The family portrait of elevated errors, then just Opus

Start with the whole catalog. End with one model and a tidy UTC range.

Status page investigated degraded performance across multiple models, listing Mythos 5, Fable 5, Opus 5, Sonnet 5, Haiku 4.5, and others. Later updates narrowed to Opus 5. Official impact window 16:11 to 18:23 UTC. Resolved same day.

outagereliabilityclaude
2026-08-17infrasev 4/5official

Opus recovered. Sonnet stayed for the second act.

Two models, one incident, staggered dignity.

Status page investigated elevated errors on Claude Opus 5, then reported Opus operational while Claude Sonnet 5 stayed degraded. Fix implemented and incident resolved the same afternoon UTC. Affected claude.ai, API, Claude Code, and Claude Cowork.

outagereliabilityclaude