Claude outage and Opus glitches

Claude users were hit by two separate problems on April 13: a service outage that blocked logins, and a fresh wave of complaints that Opus 4.6 had become slower and less reliable. (anthropic.statuspage.io) (venturebeat.com) Anthropic’s public status page said elevated login errors affected Claude.ai and Claude Code from 15:31 to 16:19 Coordinated Universal Time on April 13, with investigation starting at 15:40 UTC and the issue marked identified at 16:35 UTC. On April 14, the same status page also showed a separate unresolved incident on usage and analytics admin application programming interface endpoints. (anthropic.statuspage.io) Downdetector’s Claude page showed a spike in user reports during the disruption, and its category breakdown listed Claude Code as 70% of reported problems, the Claude application programming interface at 16%, and login at 7%. Third-party outage trackers later marked the April 13 incident resolved. (downdetector.com) (statusgator.com) Claude Opus 4.6 is Anthropic’s flagship model for hard coding and long, multi-step work. When Anthropic launched it on February 5, the company said it improved coding, debugging, long-context work, and “adaptive thinking,” a setting that lets the model decide how much reasoning effort to spend. (anthropic.com) That same adaptive-thinking design is now part of the dispute. VentureBeat reported that developers on GitHub, X, and Reddit have accused Anthropic of making Opus 4.6 and Claude Code less capable in recent weeks, while Anthropic employees publicly denied degrading models to manage capacity. (venturebeat.com) One of the most-circulated complaints came from a GitHub issue filed April 2 by Stella Laurenzo, whose LinkedIn profile identifies her as a senior director in Advanced Micro Devices’ artificial intelligence group. VentureBeat said her analysis covered 6,852 Claude Code session files, 17,871 thinking blocks, and 234,760 tool calls, and argued reasoning depth had fallen since February. (venturebeat.com) A separate viral benchmark claim said Opus 4.6 had fallen from 83.3% accuracy to 68.3% on a hallucination test. Yahoo Tech reported critics called that comparison flawed because the earlier score used six tasks, the retest used 30 tasks, and the six overlapping tasks were nearly unchanged at 87.6% versus 85.4%. (tech.yahoo.com) Anthropic’s own Opus 4.6 system card shows the company has revised published benchmark numbers before, including February and March updates tied to improved cheating-detection pipelines on Humanity’s Last Exam and BrowseComp. Those edits do not document a public Opus 4.7 release, but they do show Anthropic has been adjusting how it measures and reports model performance. (anthropic.com) As of April 14, Anthropic’s official channels still list Opus 4.6 as the current flagship public Opus model. The outage is over, the complaints about quality are not, and Anthropic has not publicly announced an Opus 4.7 launch. (anthropic.com) (anthropic.statuspage.io)