← Back to all guides

Debug API Errors: Fix 401, 402, 429 and 5xx

Failed API requests are easier to fix when you treat the HTTP status code as a decision point rather than a generic error. A 401 usually means authentication is wrong, a 402 points to payment or account limits, a 429 signals rate limiting, and a 5xx response usually indicates a temporary server or upstream problem.

Before changing code, record the request timestamp, HTTP method, URL, status code, response body, relevant response headers, model name, and a provider request ID if one is returned. Remove API keys and personal data from logs. This evidence prevents you from debugging the wrong layer.

Start With a Minimal Request

Reproduce the failure with one small request using a known model and a low token limit. Test the same credentials outside your application with curl or a short SDK script. If the minimal request fails, the problem is probably configuration, billing, access, or the relay service. If it succeeds, inspect your application’s concurrency, prompt size, timeout, and retry behavior.

Check the endpoint carefully. The base URL, API path, authentication format, and model identifier must match the SDK or tool you are using. For 59API, the API base URL is https://api.59api.com. When configuring an OpenAI-compatible SDK, set its base URL through the SDK’s supported option instead of hard-coding a second endpoint elsewhere in your application.

401: Authentication Failed

Confirm that the key is present in the process that is actually making the request. A key exported in your terminal may not be available to a Docker container, CI runner, IDE extension, or system service. Print only whether the variable exists, never the key itself.

• Check the authorization scheme, usually Bearer followed by one space and the key.
• Look for expired, revoked, truncated, or incorrectly scoped credentials.
• Verify that your code is not sending an old environment variable or a placeholder value.
• Compare the failing request with a successful request, including the base URL and model name.

Do not repeatedly retry a 401. Retries cannot repair invalid credentials and may make logs harder to read.

402: Payment or Account Limit

A 402 response normally requires an account action, not a code change. Check your balance, payment status, spending limit, and whether the selected model is enabled for your account. Also confirm that a long prompt or unusually large output has not exhausted available credit.

This is where a pay-as-you-go relay can reduce operating cost. 59API provides low-cost access to Claude Opus, Sonnet, Haiku, and Fable plus GPT models, while using native official-quality models rather than a downgraded substitute. Review pricing and balance before switching models or adding automatic retries. A referral rebate can also reduce effective cost for eligible accounts.

429: Rate Limited

First inspect the Retry-After header. If it is present, honor it. Otherwise use exponential backoff with jitter, such as approximately 1, 2, 4, and 8 seconds, with a maximum delay. Limit total attempts and cap concurrent requests with a queue or semaphore.

• Reduce parallel requests from workers, browser tabs, or background jobs.
• Reuse connections and avoid polling more often than necessary.
• Shorten prompts and output limits when token-based limits are involved.
• Retry only requests that are safe to repeat, and respect the service’s published limits.

Track request rate, token volume, and queue depth. A 429 that appears only during traffic spikes is usually a capacity-planning issue, not an invalid API key.

5xx: Server or Upstream Failure

For 500, 502, 503, and 504 responses, check whether the issue is isolated to one model, endpoint, region, or time period. Retry transient failures with exponential backoff, a strict attempt limit, and a request timeout. Use an idempotency strategy for operations where repeating a request could cause side effects.

Do not retry forever. After two or three failed attempts, return a useful error to the caller, preserve the request ID, and alert the operator. If failures persist, test a smaller request and another available model, then check the relay’s service status or contact support with timestamps and sanitized request details.

Quick Debugging Checklist

• Capture status, body, headers, model, endpoint, and request ID.
• Reproduce with one minimal request.
• For 401, verify key loading, Bearer formatting, and endpoint configuration.
• For 402, check balance, limits, payment status, and model access.
• For 429, honor Retry-After and control concurrency.
• For 5xx, retry briefly with backoff, then investigate service health.
• Keep secrets out of logs and add monitoring for status-code trends.

Developers who want Claude and GPT access without committing to a large fixed plan can try 59API at https://api.59api.com. Its pay-as-you-go model, compatibility with Claude Code, Codex, and OpenAI SDKs, and low-cost access make it a practical option for testing and production workloads. Sign up, configure the base URL, and verify a minimal request before migrating larger traffic.