Is Claude Code Down? Live Status Checks & 529 Error Fixes

Claude Code down right now? Check the official status page in 30 seconds, run the local-vs-outage checklist, and fix 529 overloaded errors for good.

2026年10月11日

Your agent session dies mid-edit with API Error: 529, the terminal spins, and the question is always the same: is Claude Code down, or is it me? Guessing wrong costs real time — restarting, re-logging, nuking your config while Anthropic is actually having a bad afternoon, or sitting out an outage that was one settings fix away.

This is the checklist we run, in order, before touching anything. It was written on October 11, 2026, in the middle of a noisy week: overloaded (529) errors have been spiking, and Claude Code 2.1.296 shipped on October 9 with a new escape hatch for exactly that error. Everything about the status page below was verified against status.claude.com on that date — and because status is the most perishable topic on this site, each section says what is a durable fact and what expires.

Quick answer: the 30-second official check

Open status.claude.com — that is the official, complete status page. Not a third-party "is it down" aggregator, not Reddit. If Anthropic is having a platform-wide incident, it is on that page within minutes of the first pager firing.

What you will find there, as verified on October 11, 2026:

Six monitored components, each with its own status light: claude.ai, Claude Console (platform.claude.com), Claude API (api.anthropic.com), Claude Code, Claude Cowork, and Claude for Government. Read the Claude Code row first; if you hit errors through the API or an app built on it, check the Claude API row too.
A five-level status legend: Operational, Degraded Performance, Partial Outage, Major Outage, Maintenance. "Degraded Performance" is the one people scroll past — it often explains "it works but everything is slow and erroring sometimes" better than a flat green page.
Active incidents with lifecycle states: Investigating → Identified → Monitoring → Resolved, with timestamped updates. The state tells you where the on-call team is; "Monitoring" means a fix is in and they are watching it settle.
A Past Incidents feed covering the last 90 days, day by day, with per-incident write-ups — plus RSS and Atom feeds and subscribe options (email, SMS, Slack, Microsoft Teams, webhook) if you want paged when it breaks.

One structural thing to internalize before you trust the page: a green status page does not mean zero 529s. The status page tracks platform-level incidents. The 529 "overloaded" error is a per-model capacity signal — one model can be drowning while the page stays green. When we checked, the incident history showed exactly this shape: "Elevated errors for Claude Opus 5.5" logged on October 6, and "Elevated errors for Claude Mythos 5.1 and Claude Fable 5.1" on October 5 — model-scoped events, while Claude Code itself stayed Operational through most of that window.

Timeliness note: component list, legend, and incident examples above were checked on 2026-10-11. Status pages change by the minute — treat the live page as the truth, and this article as the manual for reading it.

Is it Anthropic or is it you? The five-step self-check

Run these in order. Steps 1 and 2 settle the "down" question; steps 3 through 5 catch the local causes that masquerade as outages.

1. Check the component, not the homepage. On status.claude.com, find the Claude Code row. Green there while you still see errors? You are almost certainly dealing with a local problem or a per-model capacity issue, not an outage.

2. Read the actual error, not the vibe. The HTTP-style code in the error message is the diagnosis:

Error	What it means	Your move
`529` overloaded	Anthropic capacity is full	Retry with backoff — see the next section
`429` rate limit	Your account hit a limit	Check plan usage, slow the loop
`401` / permission	Auth problem	Re-login, check `ANTHROPIC_API_KEY`
`5xx` / connection refused	Network path broken	Proxy, VPN, DNS — or a real incident

3. Check your version. Run claude --version. If it is not current, run claude update — a surprising share of "Claude Code is down" reports are a fixed bug in an old client. As of publication the current release is 2.1.296 (October 9, 2026), and it carries three markers that prove you are on it: the new CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS environment variable (more below), the MCP server description limit raised from 2048 to 4096 characters, and Sonnet 5.5 cache reads billing at $0.10 per MTok instead of $0.20.

4. Check the network path. Corporate proxies, VPNs, and DNS filters quietly break WebSocket and streaming connections in ways that look exactly like an outage — intermittent hangs, connection resets, half-finished responses. If you are on a managed machine, try one session from a network without the middleboxes before concluding anything.

5. Check quota before you check Twitter. A exhausted plan limit or an out-of-credit API key produces errors that users frequently misreport as downtime. Look at your rate-limit and usage pages in the console, and if the metering itself is the confusing part, our Claude Code usage limits guide decodes it.

If you have been through all five and it is still broken with a green status page, the generic troubleshooting ladder in Claude Code stuck: troubleshooting picks up where this checklist ends.

The 529 "Overloaded" error — this week's event

Most of what gets posted as "Claude Code is down" this month is not an outage. It is the 529 overloaded error: Anthropic's servers telling you, honestly, that the model you asked for has no free capacity for you right now. It is not a ban, not a corrupted install, and usually not even a billing issue. It is congestion — and it concentrates at predictable times (more on that in the next section).

The reason it is in the news this week: Claude Code 2.1.296, released October 9, 2026, added a knob for it. From the official changelog, verbatim:

Added CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS environment variable to set a longer maximum delay for the backoff when retrying an overloaded (529) request.

What that means in practice: when a 529 hits, Claude Code retries automatically with exponential backoff, and the old ceiling on that backoff could arrive before Anthropic's capacity did — so long-running agent sessions died mid-task after a few minutes of congestion. This variable raises that ceiling. Set it in milliseconds, for example:

export CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS=300000

would let the backoff stretch to five minutes between retries instead of giving up early. Put it in your shell profile if you run unattended agents — a batch job that rides out a fifteen-minute congestion window is a batch job that finishes. Note we are deliberately not quoting a default value here: the changelog documents the variable, not its factory setting, and we would rather under-claim than print a number we have not verified.

Two practical notes. First, the env var treats the symptom patiently; it does not create capacity. At genuine platform saturation, every client is retrying and the fix is time. Second, if you want the full playbook — 529 versus 500 versus 429, wait-window strategy, model downshifting mid-session — our dedicated Claude Code overloaded (529) guide goes deeper than this summary.

Why is Claude Code down? The recurring patterns

"Why is Claude Code down" trends for a reason: the downtime is not random. Three patterns explain most of it. What follows is a pattern read from the status page's own 90-day incident history as it stood on October 11, 2026 — we do not keep a private incident archive, so every example below is one the official page itself logged.

Pattern 1: capacity overload at US peak hours. Anthropic's user base concentrates in North American working hours, and so does the load. If you are elsewhere on the planet — Asia, Europe mornings — you are structurally using the service during its calmer shifts. Chinese-speaking users get this for free by timezone: Beijing evenings (21:00 onward) overlap the US East Coast workday, while Beijing working hours land in the American middle of the night. The 529 error concentrates in the former window.

Pattern 2: launch-day congestion. New models and major releases pull a wave of demand into the exact hours after announcement. The 5.5 generation launched in a burst — Opus 5.5 on September 22, Sonnet 5.5 on September 28, Haiku 5.5 on October 7 — and the same week's incident log carries "Elevated errors for Claude Opus 5.5" (October 6) and the paired Mythos/Fable event (October 5). We are not claiming each incident was caused by its launch; we are noting that congested weeks and launch weeks have been the same weeks.

Pattern 3: model-scoped incidents that spare the status light. As covered above, incidents can be scoped to one model while the Claude Code component stays green. If errors name a specific model, the fastest check is whether the model name appears in recent incident entries — and the second-fastest is downshifting to a different model for the rest of the session, which sidesteps the question entirely.

The honest summary: Claude Code's worst hours are predictably the US workday, launch windows, and whatever one model happens to be having a bad day — and none of those require the status page to turn red.

What to do while it's down

Once you have confirmed congestion rather than a local fault, three moves, in ascending order of commitment.

Shift your hours, not your expectations. If your deadline allows it, running heavy agent work outside the US 9-to-5 measurably cuts 529 exposure. For off-hours batch runs, the retry-backoff variable from the last section is the difference between riding out a congestion spike and dying on the first one.

Downshift the model tier. Congestion is usually per-model. If Opus or Sonnet is the one throwing 529s, dropping the session to Haiku for the mechanical parts — refactors you can review, tests you can rerun — keeps you productive, and Haiku 5.5 is a far better fallback than any previous Haiku. The full tier-by-tier economics are in is Claude Code pricing worth it.

Take the free route when the bill or the outage both hurt. The more structural answer: Claude Code is a harness, and the harness can point somewhere else. DeepSeek exposes an Anthropic-compatible endpoint, so the same CLI, commands, and muscle memory run against different models with per-token billing — deepseek-flash starts at $0.15 per million input tokens off-peak. You are trading frontier-model quality for a bill you control and immunity from Anthropic's capacity events; those are different models and this is a real trade, not a free lunch. If the outage-plus-529 week has you curious, the two-variable wiring is in use DeepSeek in Claude Code.

One adjacent distinction worth a paragraph: slow is not down. If sessions crawl but complete, you are looking at throttling, network distance, or a long-context tax rather than an outage — and the fixes live in Claude Code slow: causes and fixes, not on this page.

Why is Claude Code down — FAQ

Why is Claude Code down right now?

The current-answer machine is status.claude.com — check the Claude Code component row and the active incidents list first. In practice, when it is down "right now" it is usually one of three things: per-model capacity congestion showing up as 529 errors at US peak hours, a model-scoped incident (where the page may still show green for Claude Code itself), or maintenance. If the page is fully green and you still cannot connect, work the five-step self-check above — version, network path, quota — because the fault is then almost certainly on your side of the wire.

How do I know if Claude Code is down for everyone or just me?

Two signals settle it. The status page's Claude Code row turning yellow or red means everyone — that is the "for everyone" case. For the "just me" case, the tell is error shape: auth errors (401), rate-limit errors (429), and connection-level failures that persist across networks are local, while a 529 overloaded error means Anthropic's capacity is full regardless of who you are — shared pain, even though the page may be green. Cross-checking a community source like the Claude Code subreddit helps confirm shared pain, but the official page remains the only authoritative source.

What does the 529 overloaded error in Claude Code mean?

It means the model you requested has no free capacity at that moment — a congestion signal, not an account problem. Claude Code retries automatically with exponential backoff. Since version 2.1.296 (October 9, 2026) you can raise the maximum backoff delay with the CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS environment variable, which lets unattended sessions survive longer congestion windows instead of failing out early.

Does the Claude status page show 529 overloaded errors?

Often, no — and this is the single most misunderstood thing about the page. The status page tracks platform-level incidents per component, while 529s are per-model capacity signals that usually never rise to an incident. The exception is saturation bad enough to become an official event, like the "Elevated errors for Claude Opus 5.5" entry logged October 6, 2026. Green page plus 529 errors is a normal combination, not a contradiction.

How do I get notified when Claude Code is back up?

Subscribe on the status page itself. It offers RSS and Atom feeds plus email, SMS, Slack, Microsoft Teams, and webhook subscriptions — per-component or page-wide. For a team running agents in CI, the webhook is the practical choice: pipe it into your alerting and you will know about the next incident before your first failed run tells you.

How do I check which Claude Code version I am on?

Run claude --version; update with claude update. As of October 11, 2026 the current release is 2.1.296, and you can confirm you have it by the presence of the CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS environment variable support. If you are several versions behind, updating is the first thing to try before reporting anything — a meaningful share of "it's down" reports are already-fixed client bugs.