Common API errors from OpenAI, Anthropic, and Google with causes and fixes. Quick lookup for debugging.

Last updated: March 2026

Endpoint: https://api.anthropic.com/v1/messages

Anthropic-specific notes:

• max_tokens is required – omitting it causes a 400 error
• 529 is a non-standard HTTP code unique to Anthropic
• Acceleration limits: rapid traffic ramp-up can trigger 429 even within quota
• Streaming responses can error after returning HTTP 200 (handle SSE errors)

Endpoint: https://api.openai.com/v1/chat/completions

OpenAI-specific notes:

• Token limit errors return 400, not a separate code – check error.code field
• Content policy violations return 400 with error.code: “content_policy_violation”
• Fine-tuned model 404s often mean the model is still training or was deleted

Endpoint: https://generativelanguage.googleapis.com/v1beta/models/...

Google-specific notes:

• Free tier rate limits were reduced 50-80% in December 2025 – many 429s are solved by enabling billing
• RESOURCE_EXHAUSTED accounts for ~90% of developer-reported errors
• Safety filter blocks return 200 with empty candidates – not an HTTP error
• Use generationConfig.candidateCount: 1 to avoid unexpected billing

import time, random
 
def retry_with_backoff(func, max_retries=5, base_delay=1.0):
    # Universal retry with exponential backoff + jitter
    retryable_codes = {429, 500, 502, 503, 504, 529}
    for attempt in range(max_retries):
        try:
            return func()
        except Exception as e:
            status = getattr(e, 'status_code', getattr(e, 'status', None))
            if status not in retryable_codes or attempt == max_retries - 1:
                raise
            delay = base_delay * (2 ** attempt) + random.uniform(0, 1)
            if status == 429:
                retry_after = getattr(e, 'headers', {}).get('retry-after')
                if retry_after:
                    delay = max(delay, float(retry_after))
            time.sleep(delay)