TL;DR Open status.claude.com and check the component you use: claude.ai for the chat, Claude Code for the terminal, Claude API for your app. If it's green and still failing for you, confirm on DownDetector, then run
/doctorto rule out your install. If everything's clean and Claude still "feels dumber today", the culprit is almost always your context, not the model.
When Claude stops answering, the first question is whether it's down for everyone or just for you. You can settle it in 30 seconds, in this order: the official status page, other people's reports and, if you use Claude Code, a check of your own install.
1. status.claude.com: the first place to look
Anthropic runs a status page with one component per product. Each one can go down on its own, so find yours:
| Component | What it is |
|---|---|
| claude.ai | The web chat and the apps. Tracked separately from Claude Code |
Claude Console (platform.claude.com) |
The API console: keys, usage and billing |
Claude API (api.anthropic.com) |
The backend your prompts hit. If it's down, Claude Code is too |
| Claude Code | The CLI and its sessions |
| Claude Cowork | Cowork, in the desktop app |
| Claude for Government | The service for public sector agencies |
What each state means
| State | What's happening | What you do |
|---|---|---|
| Operational | Working normally | If it fails for you, move on to steps 2 and 3 |
| Degraded Performance | Working, but slow or with scattered errors | Retry. A long task may get cut off halfway |
| Partial Outage | Failing for some users or some models | If the incident names a model, switch to another with /model |
| Major Outage | Down for almost everyone | Wait. There's nothing to fix on your side |
| Maintenance | Scheduled maintenance | Read the notice to see what it covers |
Below the components is the incident history. Each incident says which component or model it hits: "Elevated errors for Claude Opus 5.5" is not the same as "Elevated errors on claude.ai, Claude Code, Claude Cowork and the Claude API".
2. DownDetector: to corroborate
The official page sometimes lags behind an incident, because the team has to confirm it before posting. DownDetector collects user reports in real time:
https://downdetector.com/status/claude-ai/
If DownDetector shows a spike and status.claude.com is still green, wait 10 minutes and check again. A real spike means it's going down, even if nobody has said so yet.
3. /doctor: rule out your Claude Code install
If both checks above come back clean and you use Claude Code, the problem isn't Anthropic. It's your install. Run:
/doctor
It audits installation, settings, MCP servers and context usage in one pass. If Claude Code won't even start:
claude doctor
That runs the check from outside a session. The full story is in the dedicated /doctor tip.
What to do when Claude isn't working and the page is green
Three things get mistaken for an outage:
- You've hit your usage limit. The limit notice isn't an outage: your plan is used up until the window resets. How the 5-hour and weekly windows work is in Claude Code usage limits.
- Your network. Try another connection or turn off your VPN. If that works, the problem was somewhere in between.
- An old Claude Code version. If the auto-update got stuck, you're carrying bugs that are already fixed. How to check is in how to update Claude Code.
When will Claude be back?
The page doesn't give a time, but every incident moves through phases and the phase tells you where it stands:
| Phase | What it means |
|---|---|
| Investigating | They know something's wrong and are looking for the cause |
| Identified | They have the cause and are rolling out the fix |
| Monitoring | The fix is in and they're watching it hold. It usually works by now |
| Resolved | Closed |
Between phases there may be Update entries with the latest they know. If it's in Monitoring and still failing for you, close the session and open a new one before you call anything broken.
"Claude is dumber today": what's actually happening
The model doesn't change between sessions. Anthropic ships numbered versions and /release-notes tells you if anything changed in your CLI. What does change is your context. Three symptoms that get mistaken for a degraded model:
Autocompact thrashing. Your context window fills, autocompact shrinks it, and the next tool result fills it again. Claude Code stops retrying to avoid burning tokens. If it starts right after you lowered the autocompact window, the window is too small for your context: I measured it here. Symptom: short, slow responses that feel "lazy". Fix:
/compact "keep only the plan and the diff"
# or, if the earlier conversation isn't needed:
/clear
MCP garbage. An MCP server returns a huge blob (a schema, a database dump, a Playwright result) and eats half your context. Symptom: the session slows down right after a tool call. Fix: disable that MCP with /mcp or start another branch of the conversation with /branch.
Anchoring. You're 200 turns in and the first misread poisoned everything since. Symptom: Claude repeats mistakes even when you point out the right answer. Fix: ask Claude for a summary of the plan and what's done, run /clear and paste it into the clean session.
If you did all three and nothing changes, it's probably Claude. Report it with /feedback.
Subscribe before the next outage catches you
If your work depends on Claude, don't wait to open the page when something feels off. Hit "Subscribe" on status.claude.com:
| Channel | Best for |
|---|---|
| Email / SMS | Critical outages, outside the browser |
| Slack / Microsoft Teams | A notice in your team's channel |
| Webhook | Your own dashboard or bot. It also fires on each component's status changes |
| Atom / RSS | If you keep everything in a reader |
Subscribe once and you'll know before your team.
Pairs well with
/doctor: the local diagnostic./contextcommand: where your context window is going.- How to update Claude Code: a stale version sometimes behaves oddly and gets blamed as "getting dumber".
- When the auto-update fails: why your version went stale without you noticing.
- "Premature close": when that error isn't your network but a git or gh that isn't installed.
- Slash commands cheat sheet: includes
/compact,/clearand/feedback. - Claude Code slow, hanging, or eating RAM: when it's neither an outage nor context, but the process itself.
- Why Claude gets worse in long sessions: context hygiene for when everything's green and only quality drops.
Official docs: Troubleshooting · Status page
