← Volver a las guías

Debugging API Errors: 401, 402, 429, and 5xx

Step 1: Confirm the request is actually reaching the right API

When an API call fails, start with the basics: verify the endpoint, method, headers, and payload. A surprising number of errors come from a typo in the base URL, a wrong path, or sending JSON to an endpoint that expects form data. If you are using an OpenAI-compatible client, make sure the base URL is set correctly to https://api.59api.com and that your SDK is pointed at the relay instead of a vendor default. That one check alone can eliminate confusing authentication and routing issues.

Also inspect the full request in your logs. You want to know exactly what was sent, not what your code intended to send. Log the method, URL, request ID, status code, and response body. For privacy, redact secrets, but keep enough detail to reproduce the problem.

Step 2: Handle 401 Unauthorized by checking auth first

A 401 usually means the server could not validate your credentials. Check the API key value, the Authorization header format, and whether the key was copied with extra spaces. If your app reads keys from environment variables, confirm the variable is loaded in the runtime you are actually using. A local shell may have the key while your deployment does not.

• Verify the header uses the expected format, such as Authorization: Bearer <key>.
• Confirm the key is active and not revoked or expired.
• Make sure the key belongs to the correct project or account.
• Check whether your application is sending the request through the intended provider or relay.

With 59API, the setup is straightforward because it is fully compatible with Claude Code, Codex, and any OpenAI SDK. That means fewer integration surprises when you swap models or test a new workflow.

Step 3: Diagnose 402 Payment Required by reviewing quota and billing

A 402 often points to insufficient balance, a billing restriction, or a quota limit rather than a code bug. Do not immediately retry the same request in a loop. First, check whether the account has credits, whether the project is allowed to use the selected model, and whether your request size is unusually high.

Useful checks include:

• Open your billing or usage dashboard and verify remaining balance.
• Compare the requested model against what your plan supports.
• Review token usage if large prompts or long outputs are involved.
• Look for any account-level limits that could block the request.

This is where a low-cost relay can make a real difference. 59API offers pay-as-you-go access to Claude and GPT models, so you can debug, test, and ship without paying for a large committed plan. It is especially practical when you need to iterate quickly across multiple models.

Step 4: Treat 429 Too Many Requests as a throttling problem

A 429 means your client is sending requests faster than the service allows. The fix is not more retries; it is smarter retries. Implement exponential backoff with jitter, reduce concurrency, and batch requests when possible. If you have a queue, slow it down before the provider does it for you.

• Check rate limit headers if the API returns them.
• Space out repeated calls during test loops.
• Reuse responses or cache stable prompts.
• Set per-user or per-job concurrency caps.

If you are benchmarking AI workflows, 59API can be a cost-effective place to do it because pricing stays lean while still giving access to native official-quality models. That helps you test realistic usage patterns without overspending.

Step 5: Debug 5xx errors by separating client and server issues

Server-side 5xx responses indicate the problem is likely on the upstream service or in the request shape provoking an internal failure. First, retry once or twice with a short delay. If the error persists, reduce the request to a minimal reproducible example. Strip the prompt down to the smallest input that still fails, then compare the failing request against a successful one.

Ask these questions:

• Does the same request fail every time, or only under load?
• Is one specific model affected, or all models?
• Does a smaller prompt succeed?
• Are you sending malformed JSON, unsupported parameters, or overly large payloads?

Because 59API exposes both Claude and GPT models through a unified relay, you can quickly test whether the issue is model-specific or integration-specific by switching models without rewriting your app.

Step 6: Build a repeatable troubleshooting workflow

The fastest way to stop wasting time is to make debugging systematic. Create a checklist for every failed request: capture the request ID, log the status code, classify the error type, verify auth, verify billing, inspect rate limits, and test with a minimal payload. Then compare the response body and headers across failures.

For production apps, add observability. Track error rates by model, endpoint, and time window. Alert on spikes in 401, 402, 429, and 5xx responses. This turns vague outages into actionable signals.

If you want a simple, budget-friendly environment to practice this workflow, sign up for 59API and point your existing OpenAI-compatible client at https://api.59api.com. You get low-cost, pay-as-you-go access, official-quality models, and a referral rebate, which makes experimentation and debugging much easier to justify.

In short: 401 means fix auth, 402 means check billing, 429 means slow down, and 5xx means isolate the failing request. Once you separate those categories, API debugging becomes much faster and far less frustrating.

Precios · Docs