Claude Opus 5, Fable and Mythos hit an 80-minute error spike

A quiet cycle turned operational overnight: Anthropic's flagships threw elevated API errors before recovering, while OpenCode and Copilot CLI shipped model and permission updates.

Nowline SEP 22 3:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Opus 5 was the slowest to recover

    Anthropic's status page logged elevated error rates across Claude Opus 5, Fable 5/5.1 and Mythos 5/5.1 from ~00:50 to 02:10 UTC on Sep 22. A fix was identified by 01:17 and Fable/Mythos were back by 01:35, but Opus 5 needed extra remediation and only fully cleared at 02:35. If your agents or Claude Code runs threw overloaded/5xx errors on Opus 5 overnight, it was upstream — not your code.

  • OpenCode adds DeepSeek V4.1 Flash and Grok 4.7

    OpenCode v1.18.32 (Sep 21) wired DeepSeek V4.1 Flash into the Zen provider and Grok 4.7 into Zen and Go, so both are selectable in the open-source CLI now — a cheap fast model and a fresh flagship with no config gymnastics. The release also fixed Bedrock image-attachment hoisting (now limited to Claude, Nova and Llama 4) and Together AI streaming usage reporting.

  • Copilot CLI tightens what a session can approve

    GitHub Copilot CLI v1.0.88-1 (Sep 22) now preserves /allow-all through managed-settings refresh failures and remembers exact per-path approvals without granting the parent directory — grants show in /list-dirs and clear with /reset-allowed-tools. Sandboxed network denials from proxy-tunnel failures now print bypass guidance. Fewer accidental broad grants if you run it in CI or agent loops.