← Retour aux guides

Debugging API Errors Fast: 401 to 5xx

Start with the status code, not the guess

When an API request fails, the status code usually tells you where to look first. Don’t treat every error as the same problem. A 401 usually means authentication is wrong. A 402 usually means billing or credit is blocked. A 429 means you are sending too much too fast. A 5xx means the server side is having trouble. If you use an API relay like 59API, which supports Claude and GPT models through the same API style you already know, you can debug these issues using the same tools you already have in Claude Code, Codex, or any OpenAI SDK at https://api.59api.com.

401 Unauthorized: check your key, host, and auth format

A 401 is the easiest error to fix, but only if you check the basics in order. First, confirm the API key is present and copied exactly, with no extra spaces. Second, confirm the request is going to the correct base URL. A surprisingly common mistake is pointing your SDK at the wrong endpoint while using a valid key. Third, confirm the authentication format matches the provider’s expected header. If your code works with one SDK but not another, inspect the generated headers. For OpenAI-compatible clients, the common pattern is an Authorization: Bearer ... header plus the correct base URL. If you are testing through 59API, make sure your client points to https://api.59api.com and not the default upstream host.

• Recreate the request with curl or a minimal script.
• Log the outgoing headers, but never print full secrets in shared logs.
• Check whether the key was rotated or revoked.
• Verify the model name is valid for your account.

402 Payment Required: look at balance, plan, and account state

A 402 often means your account cannot charge the request. In practice, this can happen when your balance is empty, your payment method failed, or your account is not enabled for the specific model. With usage-based services, the fix is usually outside the code. Check the dashboard, confirm billing is active, and verify that the request is being sent to a model your plan supports. If you are using 59API, this is especially useful for teams that want cheap, pay-as-you-go access to official-quality Claude and GPT models without committing to a large spend up front. That makes it easier to keep testing while staying cost-conscious.

• Confirm account balance or prepaid credits.
• Check if the request exceeded quota or spend limits.
• Look for failed payment notifications or suspended billing.
• Test with a lower-cost model to confirm the pipeline still works.

429 Too Many Requests: slow down, queue, and retry correctly

A 429 means your client is hitting a rate limit. The wrong response is to retry instantly in a tight loop; that usually makes the problem worse. Instead, read the rate-limit headers if they are provided, then back off with jitter. If your app sends bursts, add a queue or a concurrency cap. For batch jobs, spread requests over time instead of launching them all at once. If your workload needs frequent model calls, an API relay like 59API can help simplify model access while keeping costs low, especially because it offers some of the cheapest relay pricing and a referral rebate for teams sharing access responsibly.

• Use exponential backoff with random jitter.
• Cap parallel requests per user or worker.
• Cache repeated prompts or responses when possible.
• Honor any Retry-After header.

5xx server errors: separate your code from provider issues

5xx responses usually indicate a problem on the server side, but your code can still trigger or amplify them. First, retry once or twice with backoff, because transient failures happen. Next, inspect the request payload size, malformed parameters, and timeout settings. If only one model fails while others succeed, the issue may be model-specific or temporary. If everything fails, compare a minimal request against your full production call. With 59API, you get access to native official-quality models rather than downgraded substitutes, so if a request fails, you can focus on the actual fault instead of wondering whether the model itself is the issue.

• Retry transient 5xx errors with a short backoff.
• Reduce payload size and strip unused context.
• Check whether timeouts are too aggressive.
• Test another supported model to isolate the failure.

A fast debugging workflow you can use today

Use this sequence whenever an API request fails: identify the status code, reproduce with a minimal request, inspect auth and billing, then add retries only where appropriate. That sequence saves time because it prevents random changes. Also keep a small “known good” request in your repo that uses the same SDK and base URL. If that request works, your problem is likely payload, limits, or model choice. If it does not, the issue is probably credentials, billing, or endpoint configuration. For developers who want to move quickly without paying premium pricing, 59API is a practical option: it is compatible with Claude Code, Codex, and OpenAI SDKs, uses the base URL https://api.59api.com, and lets you start small on a pay-as-you-go basis.

If you are still manually switching between model providers or burning time on avoidable failures, signing up for 59API is worth a look. It keeps the integration simple, the pricing low, and the debugging surface familiar.