Whether you are in the middle of an outage right now or planning for reliability, this page tells you exactly how to diagnose the problem, what Claude's error messages actually mean, what caused past outages, and your best options while waiting.
Most reported Claude outages are actually local issues — browser cache, VPN interference, or a single session auth problem. Run these steps in order before concluding it is Claude's fault.
A successful response (HTTP 200) confirms the API is up. An HTTP 529 means the API is overloaded. HTTP 500 or 503 indicates a server-side outage. A connection timeout means network-level routing failure between you and Anthropic.
| What You See | What It Means | Likely Cause | What To Do |
|---|---|---|---|
| “Claude is at capacity” or “We’re unable to process your request” | Anthropic’s servers are temporarily overloaded — not your personal usage limit | Peak-hour traffic surge, viral news coverage, or post-launch traffic spike | Wait 5–15 minutes and retry. Try off-peak hours (early morning, late evening US time) |
| HTTP 529 “Overloaded” (API response) | The API endpoint is at capacity and actively rejecting requests | Server capacity limit reached at the infrastructure layer | Implement exponential backoff with jitter. Start at 1 second, double up to 60 seconds maximum. Do not hammer the endpoint |
| Connection timeout / page won’t load at all | Request never reached Anthropic — or is stuck in routing | VPN routing failure, CDN outage, ISP issue, or total infrastructure outage | Try step 5 (VPN off) and step 6 (mobile app) from the diagnosis above. Check your internet on another site first |
| “Something went wrong” (generic error) | Catch-all error — could be almost anything | Auth session expired, server error, or browser JS failure | Reload the page. If that fails, log out and log back in. If still failing, try incognito mode |
| Infinite loading spinner | Claude’s interface loaded but is stuck fetching your session or a response | JS/CDN asset failed to load, session auth problem, or streaming response dropped mid-transfer | Hard refresh (Ctrl+Shift+R on Windows, Cmd+Shift+R on Mac). If that fails, clear cache and log back in |
| “Your message was too long” (on a normal-length message) | Context window bug — Claude thinks the conversation is longer than it actually is | Corrupt conversation state, usually after an interrupted response | Start a fresh conversation. If the original conversation was in a Project, it is still saved — just open a new chat and continue from there |
| “Claude is not available in your region” | Your geographic location is restricted from Anthropic’s services | Regulatory compliance restriction, not a service outage | This is not an outage. Anthropic restricts access in certain countries for legal/regulatory reasons. A VPN may help (check local laws) |
| HTTP 500 or HTTP 503 | Server-side error — Anthropic’s backend is returning an error response | Deployment gone wrong, database issue, or full server failure | Check status.anthropic.com immediately. This is a real infrastructure problem — nothing you can fix on your end. Wait for Anthropic resolution |
To give you accurate expectations, here is a timeline of notable incidents based on available data from status.anthropic.com, Downdetector, and coverage from outlets including TechRadar.
An extended degraded performance period in March 2026 affected API response times across several model variants, with Opus-class models most impacted. The incident lasted approximately 2–3 hours at degraded level before recovering. claude.ai continued to function for most users during this period but experienced elevated latency. The March incident coincided with a period of heavy developer adoption following a product announcement, suggesting a traffic-driven overload rather than an infrastructure failure. API users on paid plans received status notifications within 15 minutes of the incident being classified.
A 2–3 hour outage in December 2025 affected both claude.ai and the API. This was one of the longer single-incident downtimes in Claude’s recent history. Both the consumer product and the API were impacted simultaneously, which is atypical — most incidents affect one more than the other. The timing (US business hours peak) and symmetrical impact suggested an infrastructure-level failure rather than application-layer issue. Anthropic issued a post-incident summary noting infrastructure improvements made as a result.
As of June 24, 2026, the status page shows an elevated error rate specifically on Claude Opus 4.8 that has been active across multiple incidents in June — including a documented service disruption on June 18 (06:55–07:40 UTC) and multiple Opus-specific issues on June 16. This is a model-specific pattern, not a full platform outage. Sonnet 4.6 and Haiku 4.5 have remained more stable during this period.
| Pattern | Detail |
|---|---|
| API vs claude.ai impact | In most incidents, the API degraded but stayed functional while claude.ai went fully down. Invest in API-based access for critical workflows |
| Peak US hours correlation | Incidents disproportionately cluster around 9am–2pm US Eastern — traffic peak drives overload before auto-scaling catches up |
| Resolution time | Most incidents resolved in 30–90 minutes. Incidents lasting more than 2 hours are exceptions, not the rule |
| Status page lag | Anthropic’s status page typically acknowledges incidents 10–20 minutes after they begin. Downdetector is faster for early detection |
The most common cause. When Claude is mentioned in viral content, a major tech publication covers a new feature, or Anthropic launches a new model, usage can spike 5–10x within minutes. Auto-scaling on AWS and Google Cloud does not respond instantaneously — there is a lag of 3–10 minutes before new capacity comes online. During that window, servers are overloaded and users see “at capacity” errors.
Typical duration: 15–45 minutes. Status page shows: Degraded Performance or Partial Outage. What you can do: Wait it out. Try again every 10–15 minutes. Early morning retries (before US peak hours) usually work.
Anthropic runs on a combination of AWS and Google Cloud. When an underlying cloud availability zone or specific service component fails (database, load balancer, CDN), the failure can cascade to claude.ai or the API even though the root cause is not Anthropic’s own code. These events are rarer but typically longer-lasting because the fix depends on the cloud provider restoring service.
Typical duration: 30 minutes to 2+ hours depending on the cloud provider’s fix timeline. Status page shows: Partial or Major Outage. What you can do: Check the AWS Service Health Dashboard or Google Cloud Status page to understand scope. Nothing you can do except wait.
Model version updates, new feature rollouts, and infrastructure changes occasionally introduce unexpected behavior — broken prompts, increased latency, elevated error rates. When Anthropic detects this, they initiate a rollback. During the rollback period (typically 20–60 minutes), service can be intermittent or fully down. Post-incident reviews often show these as the root cause of the more surprising, mid-day outages that do not correlate with obvious traffic spikes.
Typical duration: 20–60 minutes. Status page shows: Investigating, then Identified, then Monitoring. What you can do: The rollback is the fix — follow status page updates.
Rare but documented across all major AI platforms. A large volume of automated requests from malicious actors can overwhelm rate-limiting infrastructure before blocks are applied. Anthropic does not publicly confirm specific DDoS events for security reasons, but the pattern (sudden onset, uneven geographic impact, rapid recovery after mitigation) is distinct from organic traffic spikes.
Typical duration: 15–45 minutes once mitigation is applied. Status page shows: Usually labeled as “elevated error rate” without specifying cause. What you can do: Nothing on your end. If you are an API user, your legitimate requests will begin succeeding once mitigations are in place.
Most people searching “is Claude down” are actually experiencing degraded performance, not a full outage. The distinction matters because your response strategy is different.
| State | What You Experience | Status Page Level | API Status | Best Action |
|---|---|---|---|---|
| Total Outage | claude.ai cannot be reached at all, or every request returns HTTP 500/503 | Major Outage | Also down or returning 500s | Switch to alternative (ChatGPT, Gemini) and set up an alert for recovery |
| Degraded Performance | Site loads but responses take 30–60s instead of 5–10s; frequent mid-conversation errors; some requests succeed while others fail | Degraded Performance | Functional but slow; elevated error rate | Shorten prompts, avoid heavy tasks, retry failed requests once |
| Partial Outage | Core chat works but specific features fail (Artifacts broken, web search unavailable, file uploads rejected) | Partial Outage | Usually unaffected | Use Claude for text-only tasks until the specific feature recovers |
| Operational (your issue) | Claude seems down but status page shows green; mobile app works fine | Operational | Operational | Run the 7-step diagnostic above — VPN and browser cache are the top suspects |
During partial outages, one often works when the other does not. The Claude iOS and Android apps route differently than the web app. If you are stuck on web, open the mobile app. If mobile is also failing, you have a genuine platform outage and need an alternative.
The Claude API is architecturally more resilient than claude.ai during partial outages. If you have an Anthropic API key, you can query the API directly even when the web interface is down — as was the case during June 23, 2026. Developers who keep a simple API test script or a fallback integration are able to continue working through most partial outages.
The most productive response to an outage is to switch tasks, set an alert, and come back when Claude recovers. Options:
| Alternative | URL | Best For (in Claude’s absence) | Gaps vs Claude |
|---|---|---|---|
| ChatGPT (GPT-4o) | chat.openai.com | Writing, coding, analysis, image generation, browsing | Shorter context window on standard tier; different writing style; no equivalent of Claude’s Artifacts for documents |
| Gemini Advanced | gemini.google.com | Research with Google search grounding, long document analysis, multimodal tasks | Less precise on complex code; different tone; requires Google account |
| Perplexity | perplexity.ai | Research, fact-checking, current events — always has live web access | Not a general-purpose assistant; limited for creative writing or long-form generation |
| OpenAI API | platform.openai.com | Developers: drop-in API fallback using GPT-4o or o3 | Different API schema, different model behavior; requires prompt adaptation for best results |
If you are building a production application on the Claude API, implement automatic failover rather than manual alternatives. A basic pattern:
This pattern ensures your users see zero downtime even during full Claude outages. The key is failing over only on 5xx/529 errors — not on 4xx errors (which indicate a request problem, not a server problem).
Go to status.anthropic.com and click “Subscribe to Updates.” Enter your email. You will receive notifications for incident start, status changes, and resolution. Free. Reliable. This is the first thing every Claude-dependent developer should set up.
Visit downdetector.com/status/claude and create a free alert. Downdetector sends an email when crowd reports spike above a threshold. Because it is crowd-sourced, it sometimes triggers 10–15 minutes before the official status page acknowledges an incident.
Both services offer free tiers with 5-minute check intervals. Set up a monitor pointing to a simple API health check (a low-token request to api.anthropic.com/v1/models using your API key). Configure SMS or Slack notifications. When the API goes down, you are notified within 5 minutes. When it recovers, you are notified again — you can resume your workload without manually checking.
Anthropic’s status page provides an RSS feed at https://status.anthropic.com/history.rss. Add this to any RSS reader (Feedly, Inoreader, or your browser) to get incident updates in your existing news feed workflow without a separate email alert.
The short answer: the API is more resilient, and the difference matters most during partial outages.
| Factor | Claude API | claude.ai Web App |
|---|---|---|
| SLA target | Higher — enterprise contracts include explicit SLA commitments | Lower — consumer product, no SLA guarantee on free or Pro |
| June 2026 outage behavior | Degraded performance (slow), stayed online | Fully unavailable for ~40 minutes |
| Infrastructure layer | Direct backend connection with retry capability | Web CDN + auth layer + React app on top of API — more failure points |
| Retry capability | Built into your code — exponential backoff means most transient failures resolve automatically | Manual refresh only |
| Business-critical recommendation | Use API with retry logic | Do not depend on for production workflows |
| Enterprise support | Dedicated support line, faster incident escalation | Standard support queue |
| Platform | 2025–2026 Uptime | Notable Incidents | Paid SLA | Status Page |
|---|---|---|---|---|
| Claude (claude.ai) | ~99.12% | June 23 2026 (major, 8K+ reports), Dec 2025 (2–3h), March 2026 (degraded) | No published SLA for Pro/consumer | status.anthropic.com |
| Claude API | ~99.5%+ | Stayed degraded-but-online during June 2026 event; Dec 2025 affected | Enterprise: explicit SLA in contract | status.anthropic.com |
| ChatGPT (chat.openai.com) | ~99.0% | Multiple incidents in 2025; Dec 2024 Christmas outage was widely covered | No SLA for Plus/consumer | status.openai.com |
| OpenAI API | ~99.4% | Multiple 2025 incidents; generally more resilient than ChatGPT UI | Enterprise: explicit SLA | status.openai.com |
| Gemini (gemini.google.com) | ~99.5% | Few major public incidents; benefits from Google’s global infrastructure scale | Workspace plans include SLA | workspace.google.com/status |
| Perplexity (perplexity.ai) | ~98.8% | Periodic slowdowns; smaller infrastructure team means slower incident response | No formal SLA for Pro | status.perplexity.ai |
Uptime figures based on 90-day trailing data from respective status pages as of June 2026. Consumer and API products may have different actual uptime. Uptime percentage alone does not capture incident frequency or severity.
Uptime, pricing, usage limits, and honest verdicts — all verified side by side.
ChatGPT vs Claude — Full Comparison →