How to Fix OpenAI and Claude API Errors: 401, 403, 429, 529 and insufficient_quota

Quick answer: Retry only temporary errors: 429 rate limits, 500, 503 and Claude’s 529, with exponential backoff that honors the Retry-After header. Do not retry 401/403 (key, organization or region problems), 400 (bad request) or billing 429s such as insufficient_quota; those need a fix in your code or account.

Which API errors should I retry?

Code	OpenAI meaning	Anthropic meaning	Retry?
400	invalid request	invalid_request_error, also returned when a spend limit is reached	No
401	invalid auth / incorrect key / not in org / IP not allowed	authentication_error	No
402	n/a	billing_error	No
403	country, region or territory not supported	permission_error	No
404	resource not found (NotFoundError in the SDKs)	not_found_error	No
413	n/a	request_too_large (Messages 32 MB, Batches 256 MB)	No, shrink it
429	rate limit, slow_down, or a billing/spend limit	rate_limit_error	Rate limit: yes. Billing: no
500	server error	api_error	Yes
503	server_is_overloaded	n/a	Yes
504	n/a	timeout_error	Yes, or use streaming
529	n/a	overloaded_error	Yes

OpenAI’s error-code table doesn’t list 404. Its Python library maps a 404 to NotFoundError, which the error guide describes as “Requested resource does not exist”, so check the model name, file or resource ID in your request (OpenAI error codes, openai-python). Anthropic’s not_found_error likewise means the endpoint path or a resource ID is wrong (Anthropic errors).

How do I fix an OpenAI 401 error?

Check, in this order: the key is copied without spaces and not revoked; the OpenAI-Organization or project header (if you send one) matches the key; and, if the project has an IP allow list, your server’s IP is on it. OpenAI’s error-code guide lists each of these as a separate 401 message, so read the message field, not just the status.

What does insufficient_quota mean, and why doesn’t retrying help?

OpenAI now returns several billing-related 429s: credit_balance_exhausted, organization_spend_limit_exceeded, project_spend_limit_exceeded and organization_usage_limit_exceeded, and the error.type can still say insufficient_quota. All of them mean money or a limit, not traffic. Add credit, raise the project’s spend limit, or wait for the monthly usage limit to reset. Your tier’s monthly cap is in our OpenAI pricing guide.

How do I handle a real rate limit (429)?

1. Read the Retry-After header when present and wait at least that long.
2. Back off exponentially with jitter: 1s, 2s, 4s, 8s, plus a random fraction.
3. If you see OpenAI’s slow_down error: above 1M TPM, OpenAI asks you to ramp traffic up by no more than 50% per 15 minutes.
4. Long term, cache prompt prefixes, batch offline work and request a higher tier.

What is Claude’s 529 error, and how is it different from 429?

As one developer put it on X, 529 “means Anthropic’s API is temporarily overloaded, so you wait a bit and retry. A 429 looks similar and isn’t. It’s about you: your rate limit” (@waldekm). That matches Anthropic’s errors page. One extra trap: when you hit your tier’s monthly spend cap, the 429 has no retry-after, because waiting minutes will not help.

Is there a retry snippet I can copy?

import random, time
RETRYABLE = {429, 500, 503, 504, 529}
BILLING = {"insufficient_quota", "credit_balance_exhausted",
           "organization_spend_limit_exceeded", "project_spend_limit_exceeded"}

def call_with_retry(send, max_tries=6):
    for attempt in range(max_tries):
        r = send()
        if r.status_code < 400:
            return r
        err = (r.json().get("error") or {}) if r.content else {}
        if r.status_code not in RETRYABLE or err.get("code") in BILLING or err.get("type") in BILLING:
            r.raise_for_status()          # fix the request or the account
        wait = float(r.headers.get("retry-after", 0)) or min(60, 2 ** attempt)
        time.sleep(wait + random.random())
    r.raise_for_status()

Tools that skip this step simply crash on the first 429 or 503; one developer built a wrapper for exactly that reason (@topjohnwu). The official Anthropic SDKs already retry twice by default.

What else should I check?

• 403 on OpenAI: the request comes from an unsupported country or region. Use a server in a supported location; do not route around it with a shared account.
• Claude 413: split large files or use the Files API.
• Long Claude requests timing out (504): switch to streaming.

Still stuck setting up keys? See how to get a Claude API key, or browse the AI API hub.