Claude Went Degraded August 18: 4,000 Reports, Code Hit First

Anthropic opened a “Degraded performance for multiple models” incident at 16:20 UTC on 18 August 2026 (23:20 WIB) after elevated errors hit Claude Mythos 5, Fable 5, Opus 5, Sonnet 5, Haiku 4.5, and other Claude models — more than 4,000 Downdetector reports by 10:08 a.m. PT.

That 4,000 figure is a floor. Downdetector only counts people who bother to file. Most of those filings named Claude Code, the terminal coding agent.

The status page listed claude.ai, the API at api.anthropic.com, Claude Code, and Claude Cowork as affected. Consumer chat, developer platform, and both agent surfaces failed together. This is the second Claude disruption in three days.

What Anthropic has confirmed

The 16:20 UTC note is the only official update in the SQ Magazine account. The company wrote that it was investigating elevated errors on requests to the named models. It did not name a cause. It did not give a restore time. It did not estimate user count.

Downdetector passed 2,000 reports at 9:46 a.m. PT and 4,000 by 10:08 a.m. PT. A Downdetector social post said user reports indicated problems since 12:39 p.m. EDT. Treat that as user-clock color, not a second official start time. Anthropic’s clock is 16:20 UTC.

Code took the visible hit first because developers notice a broken agent in the terminal faster than a slow chatbot. That does not prove the fault started in Claude Code. It proves who files a report.

The August 16 incident is context, not a cause

Two days earlier, Anthropic logged an authentication incident on the same status page. Failures began at 21:58 UTC on 16 August, were confirmed as degraded performance at 22:07 UTC, and were marked resolved at 22:40 UTC — a 42-minute window.

That earlier event spared the Claude Console and the API. This one does not. Once the API is on the incident list, every third-party product that routes through Anthropic is in the blast radius.

Anthropic has not linked the two events. SQ Magazine is explicit: elevated errors and heavy report volume are proven. A shared root cause is not. Do not invent one in Slack for your status page.

What we do not know

Three questions stay open.

What triggered failures across so many models at once? A model-specific bug usually stays inside one alias. A multi-model spike often sits in a shared gateway, a quota layer, or an auth dependency. That is a pattern, not a diagnosis.

How many users were actually affected? Four thousand reports is not four thousand unique paying seats, and it is not four million sessions. It is a public complaint count.

Is this tied to 16 August? Only Anthropic can say. Until a postmortem lands, treat them as two incidents that share a week.

Why agent teams felt this first

Claude Code and Cowork are the products that turn a model into a loop: plan, edit, run, retry. When those loops error, a desk stops. A chatbot that answers slowly is a mood. An agent that fails mid-patch is a merge queue.

If you already keep local coding agents as a backup, this is the morning that choice paid rent. Hosted models are fine until the host is the incident.

The same habit applies to fast cloud loops on Gemini 3.6 Flash: a second path, a retry with backoff, and a skill that can degrade to “draft only” when tools fail.

Write that fallback as a Skill, not as a sticky note. Trigger, input, output contract, and a line that says: if the API errors twice, stop and wait. Check the status page before you debug your own prompt for an hour.

The GitHub outage on 17 August is the same lesson on a different host. Your agent is only as up as the forge, the model, and the identity layer.

AIO, GEO, and SEO notes for this story

The snippet is Anthropic, Claude, 18 August 2026, 4,000 reports. Degraded is the official word; outage is what users felt. Official word first. 16:20 UTC is 23:20 WIB.

What teams should do while it is degraded

Watch the incident page. Subscribe there, not to a screenshot in a group chat.

Add retries with backoff on the Claude API. Cap them. A retry storm makes a degraded service worse.

Keep a fallback model for production paths. No failover removes the risk. It just keeps a support bot from going silent.

Do not declare a prompt regression until the status page is green. Yesterday’s skill did not suddenly forget how to call a tool.

FAQ

When did the Claude incident start on 18 August? Anthropic opened the investigation at 16:20 UTC (23:20 WIB). Downdetector’s user reports clustered through the morning Pacific window, passing 4,000 by 10:08 a.m. PT.

Which models and products were affected? Mythos 5, Fable 5, Opus 5, Sonnet 5, Haiku 4.5, and other Claude models. The status page listed claude.ai, the API, Claude Code, and Claude Cowork.

How bad was it? Downdetector logged more than 4,000 reports by 10:08 a.m. PT, up from over 2,000 at 9:46 a.m. PT. Most filings pointed at Claude Code. Anthropic has not published an affected-user count.

Was this the same as the 16 August outage? Unknown. The 16 August event was an authentication incident from 21:58 to 22:40 UTC and spared the API. Anthropic has not tied the two together.

Did Anthropic say what caused it? Not in the 16:20 UTC update covered here. The company said it was investigating elevated errors. Treat any root-cause thread as unconfirmed until the status page says otherwise.

Sources

• Claude Down: Anthropic Investigates Multi-Model Errors — SQ Magazine, 18 August 2026