Anthropic reports elevated errors on Claude Opus 5, Mythos 5 and other models

What happened

Anthropic is working through a service incident that raised error rates across several of its Claude models. According to the company's status page at status.claude.com, the incident — listed as "Elevated Errors for Multiple Models" — was opened for investigation at 05:06 UTC on August 24, 2026, after requests to Claude Mythos 5, Claude Fable 5, Claude Opus 5 and Claude Opus 4.8 began failing at elevated rates.

The incident page, which reached the Hacker News front page, lists four affected surfaces: claude.ai, the Claude API at api.anthropic.com, Claude Code and Claude Cowork. That puts both the consumer chat product and the developer-facing API in the blast radius, along with the coding and coworking tools built on top of them.

How the incident unfolded

The status updates trace a roughly two-and-a-half-hour arc from first investigation to partial recovery:

• 05:06 UTC — Anthropic starts investigating elevated errors on requests to Claude Mythos 5, Claude Fable 5, Claude Opus 5 and Claude Opus 4.8.
• 05:27 UTC — The company says it has identified the cause of the elevated errors on Claude Mythos 5, Claude Fable 5, Claude Opus 5 and other Claude models, and is working on a fix.
• 06:42 UTC — Anthropic reports it is still working on the issue and promises another update shortly.
• 07:47 UTC — Error rates have stabilized on Opus 5 and Fable 5, but the company says it is still working to fully restore success rates on all affected models.

The list of affected models grew between the first and second updates: what began as four named models was broadened to include "other Claude models" once the cause was identified. The status page does not say what the underlying cause was, what the errors looked like to clients, or how far success rates dropped.

What it means for users

For developers hitting the Claude API, an incident like this surfaces as failed requests rather than a hard outage — the service stays up, but a share of calls come back with errors. That can be harder on production systems than a clean downtime window, because retry logic, idempotency and backoff behaviour all come into play, and partial failures are trickier to detect than total ones.

The inclusion of Claude Code and Claude Cowork widens the impact. Long-running coding or multi-step agent tasks are sensitive to a single failed request mid-stream, since a retry does not always resume cleanly from where the work left off.

Why it matters

First, the breadth is the story. Errors spanning Opus 5, Mythos 5, Fable 5, Opus 4.8 and other models at the same time point toward infrastructure shared across the model lineup rather than a problem with any single model — although Anthropic has not confirmed the cause, so that remains an inference.

Second, recovery was only partial as of the latest update. Opus 5 and Fable 5 had stabilized by 07:47 UTC, but the company was still working to restore success rates on the remaining affected models, meaning users on those models could still see failures.

Third, it is a reminder of how concentrated the dependency has become: a single vendor incident simultaneously affected a chat app, a public API and two agent-style products. Teams building on the Claude API will want to watch the status page for the resolution update, and to treat error budgets and graceful degradation as first-class design concerns rather than optional extras.