Anthropic outage wastes 290,000 tokens

Anthropic’s Claude service had a broader outage on April 28 that made Claude.ai unreachable for some users and triggered authentication errors on the API and Claude Code. (anthropic.statuspage.io) Anthropic’s status page said the main incident began at 17:34 Coordinated Universal Time, or 1:34 p.m. Eastern, and success rates returned to normal by 18:52 UTC. The company marked the outage resolved at 19:15 UTC. (anthropic.statuspage.io) The same status history shows two earlier April 28 incidents before the main outage: elevated errors on Claude Haiku 4.5 from 11:53 to 12:44 UTC, and elevated errors on Claude Sonnet 4.5 from 13:22 to 13:39 UTC. It also logged intermittent Claude Code review failures on April 27 and billing-related Claude.ai errors that day. (anthropic.statuspage.io) For companies using Claude through hosted infrastructure, those incidents hit two different failure points at once: the consumer chat site and the authentication path for developer tools. Anthropic’s status page lists 60-day uptime of 98.41% for Claude.ai and 98.86% for the Claude API. (anthropic.statuspage.io) A separate problem circulating among developers was not downtime but cost. In a GitHub issue opened April 16, a Claude Code user said version 2.1.111 with Opus 4.7 consumed about 36,700 tokens for a fresh session on a minimal task, versus about 21,500 on v2.1.110 with Opus 4.6. (github.com) That report did not document a single 290,000-token run. It described a measured jump of roughly 1.7 times to 2 times on the reporter’s setup, with higher system-prompt and tool overhead in the newer version. (github.com) Other Claude Code users have filed broader complaints about token burn and quota drain in recent weeks. One March issue said a monitoring proxy found cache-read ratios had fallen to 4.3%, inflating token use per turn until workarounds pushed the ratio back to 89% to 99%. (github.com) Anthropic had already acknowledged Claude Code reliability and quality problems before Monday’s outage. In an April 23 engineering post, the company said it traced recent quality reports to three separate changes, said the API and inference layer were unaffected, and said all three issues were resolved by April 20 in version 2.1.116. (anthropic.com) Anthropic’s public changelog shows Claude Code moved past v2.1.111 quickly, with versions 2.1.121 and 2.1.122 released on April 28 and v2.1.123 on April 29. The changelog lists fixes for authentication loops, structured-output errors on cloud providers, and other tooling bugs. (code.claude.com) Taken together, the April 28 outage and the recent Claude Code bug reports show two separate risks for teams building on model vendors: the model can be available but expensive, or cheaper to run but temporarily unreachable. (anthropic.statuspage.io) (github.com)