Claude's third API error incident in 11 days now hits Opus 5
A ~100-minute overnight degradation also caught Fable and Mythos — the pattern now says bake in retries. Plus an open MIT NLI model and Grok 4.7 in Copilot.
Copy markdown
What happened: ~100 minutes of elevated API errors
From 00:57 to 02:35 UTC, Anthropic reported elevated error rates on the Claude API for Opus 5, Fable 5/5.1 and Mythos 5/5.1. Downdetector logged 1,600+ reports across multiple regions before success rates returned to normal. It's resolved — but for roughly an hour and a half, live requests were failing.
The pattern: third Fable/Mythos incident in 11 days
Anthropic's status page logged elevated-error incidents on Sep 11 and Sep 15 (Mythos 5.1, Fable 5.1) — and this one escalated to include flagship Opus 5. If you call Claude in production, treat degraded responses as normal: add retries with backoff and wire up a fallback model or provider before the next one lands.
Elsewhere: openjev, one tiny open model for rerank, grade and guard
AlexWortega released openjev, an MIT-licensed NLI cross-encoder built on Qwen3.5 in 0.8B, 4B and 35B-A3B sizes. A single entailment/contradiction/neutral output covers reranking, grading and content guardrails — v2 pushes ANLI-r3 to 0.63 and WANLI to 0.77. A cheap, local stand-in for an LLM judge.
Elsewhere: Grok 4.7 is now selectable in GitHub Copilot
One day after launch, xAI's Grok 4.7 is live in Copilot for agentic coding and multi-step work — pick it from the model selector. It keeps the $2/$6 pricing tier, though heavier agent loops can still push your token bill up.