Error and cost playbooks
What the error means, why it happens, and what the gateway does about it.
OpenAI 429 rate limit
HTTP 429 with rate_limit_error, often with a retry-after header, under bursty traffic or on a new project with low limits.
Anthropic 529 overloaded
HTTP 529 overloaded_error during provider-wide load spikes; retries sometimes succeed within seconds.
Prompt cache misses
Cached-input tokens stay at 0 even though the system prompt never changes.
Reasoning-token bloat on DeepSeek and o-series models
Short classification or extraction prompts cost 5-15x more than expected; output_tokens dwarf the visible answer.
Tool-schema token bloat
Agent requests carry thousands of input tokens of tool definitions on every turn; cache reads help but the schemas themselves are padded.