Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- The final record for September 22 incident 7g1qpkyz5gxh defines the impact window as 00:50–02:10 UTC, or 08:50–10:10 Beijing time. The Resolved update was published at 02:35 UTC.
- The investigation named Mythos 5.1, Fable 5.1, and Opus 5. An intermediate update said Fable 5/5.1 and Mythos 5/5.1 had recovered while Opus 5 was still being addressed. Success rates later returned to normal for all affected models.
- The incident affected claude.ai, Claude API, Claude Code, and Claude Cowork. Official information provides no root-cause details, error codes, Regions, error rates, compensation, or task-completeness findings. Do not describe every failure in the window as 529.
- 529 overloaded_error means temporary API overload. It is different from 429 rate limiting, 401 key errors, 400 parameter errors, and proxy connection failures.
- Record Beijing time and UTC, model ID, error type, request-id, and whether the task writes externally, then pause concurrency and automatic retries for the same task.
- Use a fallback model only after tool, context, format, and cost acceptance checks. Recovery should not depend on one successful request: wait for official status to stabilize, validate with a minimal request without side effects, then restore traffic gradually.
Start with the September 22 final status: models recovered, but outcomes still need reconciliation
At 00:57 UTC on September 22, 2026 (08:57 Beijing time), Anthropic published an investigation update naming elevated errors for Mythos 5.1, Fable 5.1, and Opus 5 requests. The final record moved the actual impact start to 00:50 UTC and confirmed effects on claude.ai, Claude API, Claude Code, and Claude Cowork. Claude Console and Government were not listed among that incident's affected components.
At 01:35 UTC, Anthropic said Fable 5/5.1 and Mythos 5/5.1 had recovered while work on Opus 5 continued. At 02:11 UTC, it said success rates for all affected models were back to normal and entered Monitoring. At 02:35 UTC, it marked the incident Resolved, with a final impact end of 02:10 UTC. Resolved means the platform incident is closed; it does not prove that every conversation, code change, or external tool action during the window completed fully.
Do not combine September 22, September 15, and September 3 into one root cause
September 15 incident 6304r9jjhj34 named only Mythos 5.1 and Fable 5.1, with a window of 09:48:23–11:14:44 UTC. September 3 incident 461yvfrzpwtt also listed Mythos/Fable 5, Opus 5, Opus 4.8, and Opus 4.6. September 22 has its own incident, 7g1qpkyz5gxh. The three can share containment and recovery checklists, but not borrowed root causes, time windows, or impact scopes.
GitHub Copilot also recorded an upstream-provider incident affecting Fable 5.1 on September 15 with overlapping timing, but Anthropic and GitHub did not explicitly confirm a shared root cause. The new September 22 incident also provides no RCA. Failures through Copilot or third-party entry points should be assessed using their own status pages, model policies, and actual request evidence.
Distinguish 529, 429, 400, and network failures first
Claude API documentation defines 529 as overloaded_error, indicating temporary service overload. A 429 concerns organizational rate limits, monthly spending caps, workspace limits, or acceleration limits, while 400 usually concerns request format, content, or configuration. The September 22 incident page only says elevated errors, so record a failure as 529 only when the actual response returns 529. DNS resolution failures, TCP timeouts, TLS certificate errors, and 407 belong to connection or proxy layers. Not every failure is a platform outage.
Minimum evidence should include occurrence time and time zone, entry point, model ID, HTTP status, error.type, request-id, whether streaming was used, client and SDK versions, and whether the task already caused external side effects. For general 4xx, 5xx, and relay-layer distinctions, see the AI relay error troubleshooting guide. Use the proxy connection troubleshooting checklist only when explicit DNS, connection, TLS, or 407 evidence exists.
Six containment steps after a failure
First, pause new concurrent runs of the same task and automatic queue expansion. Second, preserve the original error and request-id. Third, check whether the official incident covers the current entry point and model. Fourth, establish whether the failed request sent email, committed code, updated tickets, or called another write interface. Fifth, separate unfinished, completed, and unknown-outcome tasks. Sixth, perform bounded validation only with minimal requests confirmed to have no side effects.
If Claude Code or Cowork disconnects mid-run, do not immediately copy the original prompt into several new sessions. Inspect the workspace, Git diff, external-tool logs, and remote systems first to determine whether the previous run partly executed. Manually verify writes with unknown outcomes before deciding to rerun. This prevents duplicate commits, tickets, messages, or deployments.
API retries: cap attempts, back off, and retain request IDs
Official Anthropic SDKs use exponential backoff for transient failures such as connection errors, rate limits, and 5xx responses, retrying twice by default and honoring retry-after when present. Wrapping that in an unlimited application loop can turn one overload into more queued requests. Set a maximum attempt count per task, total wait, concurrency cap, and circuit-breaker window.
Record each attempt's new request-id and final status, but do not place API keys, full prompts, or client data in public logs. Stop and wait immediately if a bounded sequence still returns 529, official status remains partial outage, or each retry may trigger an external write. Do not attempt to bypass capacity problems by increasing concurrency, rotating keys, or frequently changing egress.
When to use a fallback model, and when waiting is the only suitable choice
Enable a fallback model only if its context length, tool schema, thinking block, output format, latency, and costs were already validated during normal operation. A model recovering on the status page does not mean it can take over a long Opus or Fable task without loss. Test a sample without side effects first, then shift a small proportion of traffic.
Pause if the task depends on specific model behavior, is already deep into a long conversation, requires strict tool calling, or the fallback model is also within the current incident's scope. When falling back, log the original model, replacement model, switch time, and task batch, and retain a switch back. If model lists or migration configuration are unclear, consult the Claude Fable 5.1 migration and 400 error guide first.
Recovery validation: stable official status, minimal probes, and gradual traffic restoration
Recovery needs at least three kinds of evidence: the official incident reaches resolved or affected components return to normal; repeated minimal requests without side effects succeed on a fixed model; and the application's error rate, latency, and queue depth return to baseline. One success or a community report that “it works” is not enough to restore all scheduled tasks.
First disable compensating concurrency and remove duplicate queue entries, then restore read-only tasks with a small volume of traffic. Resume Agent workflows that may write to external systems afterward. Reconcile each task with an unknown outcome from the incident, checking for duplicate submissions, missed writes, and abnormal costs. If errors rise again, use the same circuit-breaker threshold to return to waiting.
When to stop troubleshooting and hand off to the platform or internal owner
When the official status page explicitly covers the current entry point and model, stop changing proxies, certificates, keys, or business prompts. For persistent 529 errors, reproducible request IDs, or organizational failures after official Resolved status, send Anthropic support redacted times, models, entry points, request-id values, SDK versions, and a minimal request. For Claude Platform on AWS, also retain x-amzn-requestid.
A proxy only affects the network path. It cannot fix upstream overload, model capacity, or official component failures. For stable access to official consoles and documentation, visit the PuppyIP website to learn about fixed network egress. Follow the service provider's Region, account, rate, and usage rules, and do not describe a network product as a solution to platform incidents.
Sources
- Anthropic official incident: September 22, 2026 elevated errors for multiple models
- Anthropic official incident: Mythos/Fable 5.1 intermittent error spikes
- Anthropic official Status API: incident 6304r9jjhj34
- Anthropic official incident: Elevated errors for multiple models
- Anthropic official: Claude API errors
- Anthropic official: Claude service tiers
Frequently Asked Questions
Does a Claude API 529 mean my key is invalid?
Usually not. Official documentation defines 529 as temporary API overload. Key issues more commonly return 401, and permission issues often return 403. Still retain request-id and check the status page.
What is the difference between 529 and 429?
529 means temporary platform overload. 429 may concern organizational rate limits, spending caps, workspace limits, or acceleration limits. Read error types and response headers for both instead of relying only on the client notice.
How many times does the official SDK retry automatically?
Claude API documentation says official SDKs use two exponential-backoff retries by default for transient failures and honor retry-after when present. Applications should still set their own total-attempt, timeout, and circuit-breaker boundaries.
Can I immediately restart a Claude Code task after a 529?
Not yet. Inspect Git, files, and external tools to determine whether the previous task partly executed. Restarting writes with unknown outcomes may duplicate commits or external actions.
Has the September 22 Mythos, Fable, and Opus 5 incident recovered?
Yes. Anthropic sets the official impact window at 08:50–10:10 Beijing time and published the Resolved update at 10:35. Tasks with unknown outcomes during the incident still need individual reconciliation.
Can I temporarily switch to an unaffected Claude model?
Yes, provided the fallback has passed tool, context, format, and cost acceptance checks and the status page shows it healthy at the time. Pause instead of switching blindly if compatibility is uncertain.
Can changing proxies or IP address fix Claude 529?
No. 529 indicates server-side overload. Investigate the proxy path only with explicit network-layer evidence such as DNS, connection, TLS, or 407 errors.