LLM API Error Codes Explained: 401, 402, 429, 500 & Fixes

LLM APIs return standard HTTP error codes, and each points to a specific cause and fix. Knowing them turns a cryptic failure into a two-minute resolution: 401 is auth, 402 is billing, 429 is rate limits, 5xx is the provider. This guide is a reference for every common code you'll see — what triggers it, how to fix it, how to handle them in code, and how failover hides the ones that aren't your fault.

The error-code reference

Every common code, its cause, and the fix:

Client errors (your side — fix, don't retry)

• 401 Unauthorized — the #1 first-call error. The key is missing, wrong, rotated, or doesn't match the endpoint's region. Re-check how the key is loaded and which base URL you're calling.
• 402 Payment Required — a billing block, not a bug. Add credits, fix the card, or check a spend cap. Your request was valid; it just couldn't be paid for.
• 400 / 422 — a malformed (400) or semantically invalid (422) request: wrong model id, bad JSON, an out-of-range parameter. Retrying won't help; correct the request.
• 403 / 404 — no access to that model/feature (403) or a wrong endpoint/model id (404). Request access or fix the identifier.

These all share one rule: don't retry — retrying an unfixed client error just fails again.

Server errors (provider side — retry, then failover)

• 429 Too Many Requests — a rate limit. Back off and retry with jitter, or fail over. See the rate-limit guide.
• 500 / 503 — the provider is faulting or overloaded. Not your fault and usually transient; retry with backoff, and fail over to an equivalent model if it persists.

Handling them in code

Branch by code — retry the transient ones, fail fast on the rest:

from openai import OpenAI, RateLimitError, APIStatusError
import time, random

client = OpenAI(base_url="https://www.datallmlab.com/v1", api_key="$DATALLMLAB_API_KEY")

def call(messages, model="openai/gpt-5.4", tries=5):
    for i in range(tries):
        try:
            return client.chat.completions.create(model=model, messages=messages)
        except RateLimitError:                          # 429 — transient, back off
            time.sleep((2 ** i) + random.random())
        except APIStatusError as e:
            if e.status_code in (500, 503):           # provider fault — retry
                time.sleep((2 ** i) + random.random())
            else:                                      # 400/401/402/403/404/422 — fix, don't retry
                raise
    raise RuntimeError("exhausted retries")

Failover hides the transient ones

The not-your-fault errors — 429, 500, 503 — are exactly the ones a gateway can absorb. On any of them it retries the request on an equivalent model from another provider, so the user never sees the error. Client errors (400/401/402/422) still need your fix, but the transient majority of production incidents become invisible reroutes. That's the difference between an LLM feature that flakes during provider hiccups and one that just keeps working.

FAQ