Claude Outage on 20 August Hits App, API, Code, and Cowork

Anthropic opened a Claude outage notice at 19:16 UTC on 20 August 2026 (02:16 WIB on 21 August) after elevated errors on some Claude models. Unite.AI, citing the Anthropic status page, listed claude.ai, the API at api.anthropic.com, Claude Code, and Claude Cowork as partial outages.

What the 20 August Claude outage reached

The affected list is the production path: consumer chat, the developer API, the terminal coding agent, and Cowork. All four were in partial outage. Claude Console at platform.claude.com and Claude for Government stayed operational. That split is the useful diagnostic. Account and admin surfaces were up. Model-serving surfaces were not.

Anthropic’s notice said it was investigating elevated errors on requests to some Claude models. The write-up used here does not name the models, a cause, or a restore time. Do not invent them. “Some Claude models” is the official subject.

Ninety-day uptime figures on the same status page still looked high: claude.ai 99.38 percent, the API 99.48 percent, Claude Code 99.4 percent, Claude Cowork 99.49 percent. Those are rolling windows, not a claim that 20 August was fine. A partial outage can sit inside a 99-point number. Report both.

The 18 August errors are context, not a cause

Two days earlier, elevated errors hit Mythos, Fable, Opus, and other Claude models from 16:11 to 18:23 UTC on 18 August 2026 (23:11 WIB on 18 August to 01:23 WIB on 19 August). Promptcrates already covered that 18 August degraded stretch. Anthropic has not, in the facts used here, tied 18 August to 20 August. Keep them as two incidents that share a week.

The same week also included a GitHub outage on 17 August. An agent that needs both the model and the forge is only as up as the weaker host. That is the operations lesson. It is not a root-cause claim.

How to write the skill so a partial outage is a stop, not a prompt bug

When claude.ai, the API, Claude Code, and Cowork are all yellow, stop debugging yesterday’s skill. Check the status page. Cap retries. Fail over if you have a second model. If you do not, degrade to “draft only” and wait.

Code and Cowork fail differently than chat. A long agent loop that errors mid-patch leaves a dirty tree. A chat that 500s leaves an empty bubble. Write the stop for the loop: two API errors, then halt, then tell the user the vendor is the incident.

Do not declare a prompt regression while the status page still says partial outage. Do not invent a shared cause with 18 August. Do not flatten 99.38 percent into “Claude is unreliable this quarter” unless you are quoting the page and the window. Watch the incident page itself, not a screenshot in a group chat, and keep a second model ready if Code or Cowork is in the production path.

If Code or Cowork died mid-task, recover the working tree before you retry the same skill. The citation sentence is Claude outage, 20 August 2026, 19:16 UTC, Unite.AI citing Anthropic status. Affected: claude.ai, api.anthropic.com, Claude Code, Claude Cowork, partial. Operational: Claude Console, Claude for Government. Ninety-day uptime: 99.38, 99.48, 99.4, 99.49. Also 18 August, 16:11–18:23 UTC, Mythos, Fable, Opus, and others.

Sources

• Claude Outage Hits Claude.ai, API, Claude Code and Cowork as Errors Spread Across Models — Unite.AI, 20 August 2026