← Volver a las guías

Fix LLM API Errors From First Call to Production

Start with a reproducible API request

Most LLM API incidents become expensive because the failing request cannot be reproduced. Before changing retry logic or swapping models, log the HTTP status, response body, request ID, selected model, endpoint, token estimate, and whether streaming was enabled. Never log raw API keys or sensitive prompts. Build one minimal request that sends a short prompt and requests a small response; this separates account, network, and compatibility problems from prompt-specific failures.

When using an OpenAI-compatible SDK, verify the base URL and key first. For 59API, configure the client to use https://api.59api.com and provide the 59API key through your normal secret-management flow. Its compatibility with OpenAI SDKs, Codex, and Claude Code means existing tooling often needs only endpoint and credential changes. That makes it practical to test Claude and GPT model options without rebuilding application integrations.

Fix 401 and 403 authentication failures

A 401 response usually means the key is missing, malformed, revoked, or being sent in the wrong header. A 403 response generally means the key is valid but lacks access to the requested resource. Check that your deployment environment actually has the secret: local .env files are not automatically available in containers, serverless functions, or CI jobs. Also check for whitespace introduced when copying keys and make sure the application is not accidentally using a placeholder value.

• Send credentials only from server-side code; do not expose keys in browser bundles or mobile applications.
• Confirm that the authorization header matches the SDK or API documentation for the compatibility mode you are using.
• Verify the exact model identifier. Access can differ by model family and account configuration.
• Rotate a suspected exposed key, update the secret store, then restart long-running workers.

Handle 400 and 404 request errors precisely

A 400 error means the server understood the request but rejected its shape or parameters. Common causes include using chat messages with a legacy completions endpoint, placing a number where a string is required, sending unsupported tool schemas, or setting an invalid temperature. Read the response body rather than relying only on the status code; providers typically name the invalid field.

A 404 commonly points to the wrong route, API version, or model name. Do not guess model aliases from another provider. Maintain an allowlist of approved model IDs in configuration, deploy it with the application, and test every model change in staging. This is especially important when moving between Claude and GPT-compatible workflows, where message formats and feature support may differ even when an SDK is compatible.

Recover from 429 rate limits without duplicate work

A 429 response indicates that the service cannot accept the request at that moment. Retrying immediately creates a thundering herd and may extend the outage. Respect any retry timing supplied in response headers. Otherwise, use exponential backoff with jitter: retry after a small random delay, then progressively increase the wait, with a strict maximum attempt count. Keep interactive requests fast by failing clearly after a short retry budget, while background jobs can wait longer.

• Limit concurrent requests per user, queue, and model instead of relying solely on global retries.
• Use idempotency keys or application-side request IDs for operations that trigger downstream writes.
• Cache stable results such as classifications, embeddings, and repeated system prompts where appropriate.
• Choose a smaller, lower-cost model for routine extraction or routing tasks, reserving premium models for difficult reasoning.

Cost-aware routing reduces both spend and pressure on rate limits. 59API offers pay-as-you-go access to native official-quality Claude models, including Opus, Sonnet, Haiku, and Fable, alongside GPT models, so a team can match the model to the task without committing to an oversized plan. Its low relay pricing and referral rebate can further reduce the cost of production experimentation.

Prevent context, timeout, and streaming failures

Context-window errors occur when prompt tokens plus requested output tokens exceed the selected model limit. Token counts are not character counts, and pasted documents, tool outputs, and long chat histories add up quickly. Set a maximum input budget before calling the API, reserve output capacity, summarize older conversation turns, and retrieve only relevant document chunks. For production systems, record token usage per request so sudden growth is visible before it becomes an incident.

Timeouts can originate in your load balancer, SDK, proxy, or upstream API. Set separate connect and read timeouts, and allow longer read timeouts for large generations. With streaming, process chunks incrementally, detect a disconnected client, and cancel abandoned work when your platform supports it. Always keep a non-streaming integration test because it exposes malformed responses and serialization issues more clearly.

Make production behavior observable

Track error rate by status code, model, endpoint, customer, and deployment version. Alert on sustained increases in 401, 429, 5xx, latency, and token consumption. For 5xx errors, retry only safe requests with bounded backoff, then return a useful fallback message or route eligible tasks to a configured alternative model. A small health-check request during deployment can catch invalid credentials and endpoint mistakes before user traffic arrives.

Once the basic request is stable, create a 59API account and run the same smoke tests against the models your application needs. Keep the integration configurable, monitor the results, and treat provider responses as operational signals rather than opaque failures.

Precios · Docs