Claude Code overloaded (529): fix the capacity error fast
What the 529 Overloaded error really is — Anthropic-side capacity, not your account — plus the status check, the /model switch, the retry-tuning environment variable from Claude Code 2.1.296, and 529 vs 429 vs 500, shown on real screenshots.
529 at a glance
- 529 Overloaded means Anthropic's API is temporarily at capacity across all users — the official docs state it "is not your usage limit and doesn't count against your quota." It is never a ban.
- First move is always the same: check status.claude.com (or curl -I it) to see whether Claude itself is degraded — our live status guide shows you how to read it.
- Instant fixes, in order: let Claude Code's auto-retry finish, run /model and switch to a lighter tier (capacity is tracked per model), lower tool-use concurrency.
- Since Claude Code 2.1.296 you can set CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS to stretch the backoff ceiling for 529 retries in long-running sessions.
- 529 ≠ 429: a 429 is your key hitting its rate limit (your side), a 529 is everyone's capacity (Anthropic's side), a 500 is an unexpected API failure (wait a minute, resend).
How to Fix Claude API "Overloaded" Error (529 Status Code)
ChannelNexustech4:32
Claude errors reference — 529, 429, 500 & retry tuning
Docscode.claude.com/docs
Frames come from one clean screen recording — the "currently unavailable" capacity banner, the curl status check in VS Code, the model picker with the overloaded model greyed out, the raw-request 529 failover layout and the in-app support chat (Nexustech, JGlrcRDtdls). Three other videos — Extreme Guides' CLI walkthrough (_oEU-6xbcWs, commands that don't exist in Claude Code), Izoate's article screencast (fHlvs5evSi0, burned-in buttons and a channel logo) and Last Error Fixer's 400 fix (vHEi32hIh1I, burned-in captions) — supplied facts or framing but were rejected as frame sources. Error wording follows the official errors documentation and the Claude Code 2.1.296 changelog (2026-10-09). Each step deep-links to its exact second.
Video frames remain the property of their creators and are embedded here as step-by-step documentation with attribution and deep links.
Fix Claude Code overloaded (529) step by step
Step 1 — Confirm it's Anthropic, not you
- 1
Recognize what the overload state looks like
A 529 is Claude-wide capacity pressure, so its signs appear on every surface at once: the Claude app wears a "currently unavailable" banner over the composer, direct API calls come back HTTP 529 with an overloaded_error body, and Claude Code's spinner switches to "Retrying… attempt x/y". If only your terminal fails while the app works fine, suspect your local setup instead — our stuck-session guide covers that path.
The Claude app wearing the "Claude Fable 5 is currently unavailable" banner — the same capacity pressure a 529 reports to your terminal.Watch at 0:02 - 2
Check the official status page first
Open status.claude.com before touching any config — it is the single source of truth for degraded performance and outages. From a terminal, curl -I https://status.anthropic.com/ resolves straight onto the status page. A green board with 529s means the overload is model-specific and the fixes below still apply; a red board means follow the incident instead of your runbook. Our live Claude status guide explains how to read the page and which sections predict terminal errors.
curl -I against status.anthropic.com in the VS Code terminal — the 307 redirect lands on status.claude.com, the page that says whether 529s are official.Watch at 1:00 - 3
Rule out the local suspects
Run /status inside Claude Code: it shows the active model, your account and connectivity. A 529 costs you nothing — the docs are explicit that it "doesn't count against your quota" — so there is no usage line to worry about here. If /status shows the right account and a working connection, the error really is server-side; anything else (401, 403, billing errors) belongs to a different guide.
Step 2 — Immediate fixes that get you unblocked
- 4
Let the built-in retry finish its work
Claude Code already retries transient failures up to 10 times with exponential backoff before surfacing an error, and the spinner counts it down ("Retrying in 8s · attempt 2/7"). During a 529 burst the smartest move is usually to wait out those attempts instead of hammering Enter — every extra request queues behind the same capacity. If all attempts run out and the error still stands, move on to the next step.
- 5
Switch to a lighter model with /model
Capacity is tracked per model, which is why the docs' first suggestion for overload is: "Run /model and switch to a different model to keep working." Type /model in the Claude Code prompt and pick a lower tier — Sonnet- and Haiku-class models keep responding while the flagship is choked. The same picker in the Claude app marks the overloaded model "Currently unavailable", so you can see exactly which lanes are closed before you commit.
The model picker with Fable 5 marked "Currently unavailable" while Sonnet 4.6 keeps its checkmark and Haiku 4.5 stays selectable.Watch at 1:36 - 6
Lower the concurrency on your side
Parallel tool calls and background jobs multiply your share of the capacity crunch. Drop CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY to 2-3, pause other running sessions or agent teams, and avoid launching a second Claude Code instance on the same repo mid-incident. Fewer concurrent requests also means the retry loop clears faster once capacity returns.
- 7
When it outlasts the incident, contact support
If 529s keep coming after the status page has cleared, take it to Anthropic: in the Claude app, open your profile menu in the bottom-left corner and choose Get help. The in-app assistant accepts an error report directly — you'll paste the status code and when it happened, without hunting for a support email.
The profile menu with Get help right under Settings and Language — the entrance for reporting a stubborn 529.Watch at 2:20 - 8
Describe the 529 to the support assistant
In the support chat, state the status code, the model and the time window: "I've received a 529 status code on claude-opus-4-8 since 14:00". The assistant assigns a conversation ID for your records — keep it in case you need to follow up. Support can tell whether your account hit a rare persistent cluster or the whole model tier was sagging for everyone.
The support chat with "I've received a 529 status code can you help…" typed in and a conversation ID already assigned.Watch at 3:06
Step 3 — Tune the 529 backoff (Claude Code 2.1.296+)
- 9
Stretch the backoff ceiling (2.1.296+)
Claude Code 2.1.296 added CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS — the changelog's exact words: "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." Export it before launching, e.g. CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS=30000 for up to 30 s between attempts. Anthropic has not published the built-in default, so treat any specific number you find quoted elsewhere as unverified — what the variable guarantees is that long retry sequences wait longer instead of giving up early.
- 10
Pair it with the other retry knobs
Two neighbouring variables complete the setup: CLAUDE_CODE_MAX_RETRIES (default 10) raises how many attempts a transient error gets, and CLAUDE_CODE_RETRY_WATCHDOG=1 keeps retrying 429/529 errors indefinitely — the docs recommend it for unattended sessions where an early abort wastes the whole run. Watchdog plus a generous max delay turns a 529 burst into a pause instead of a crash.
Step 4 — Prevent the next overload stall
- 11
Prevent the next stall: off-peak, split work, failover
Three habits cut 529 exposure: run heavy jobs outside US peak hours (roughly 11:00-20:00 PT on weekdays), split monolithic prompts into smaller tasks — a git worktree per task keeps sessions independent — and if you call the Messages API directly, code a failover: catch the 529, hot-swap to a lighter model, retry with exponential backoff. For plan-side ceilings see our usage-limits guide; 529 itself never draws down your quota.
A raw request failover layout: the primary target is claude-sonnet-4-5, and a 529 response hot-swaps to the claude-haiku-4-5 fallback.Watch at 2:08
529 vs 429 vs 500: which error is which
Claude's errors travel with numbers, and the number decides whose problem it is. The official error reference splits them cleanly — keep this list in mind before you start deleting config files.
- 1429 rate limit — your key, project or plan hit a limit configured on your side. Run /status to confirm the active credential, check your provider console, and lower CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY. Short throttles are auto-retried; a gateway spend-limit 429 is not.
- 2529 Overloaded — the API is temporarily at capacity across all users. Server-side by definition: nothing in your code, key or config causes it, and it never counts against your quota. Wait, switch models, or stretch the backoff.
- 3500 Internal server error — an unexpected failure inside the API, usually temporary. Wait a minute and resend the same message; for a long prompt, typing "try again" resumes from where it stopped.
- 4"Server is temporarily limiting requests" — a short-lived throttle that is unrelated to your plan quota. Wait briefly and try again; no config change will lift it faster.
- 5The 60-second triage: green status page + 529 means wait it out; green status page + 429 means check your own limits with /status; red status page means anything goes — follow the incident instead of retrying blindly into it.
- 6When to actually worry: 529s clustered on one model for hours while the status page stays green deserve a support report with timestamps (step 8). Everything else is transient by design — the official docs treat all three codes as retry-worthy, not as account problems.
Same division of labor on this site: our stuck-session guide owns freezes and hangs, this page owns 529/500 capacity errors, and the usage-limits guide owns 429-style ceilings. If your session froze rather than errored, switch guides — the fixes don't overlap.