AI API errors: find it, fix it, pick a better model

Showing 15 of 15 errors

Quota and rate limits: the two 429s that need opposite fixes

OpenAI sends both rate limits and exhausted billing as HTTP 429, and confusing them wastes hours: backoff cures the first and does nothing for the second. Anthropic adds a wrinkle worth knowing — acceleration limits that fire when usage ramps too sharply, even below your ceiling. If limits keep binding, the practical escape is routing bulk traffic to cheaper, higher-throughput tiers: compare what that costs in the calculator against the current rankings.

Model not found: 2026's fastest-growing error

Three providers are retiring model lines this year — Claude Sonnet 4 and Opus 4 went dark June 15, Gemini's 2.5 line ends October 16, and OpenAI retires nine IDs on October 23. Every 404 in this category (OpenAI, Anthropic, Gemini) links into the deprecations record, where each retirement carries its replacement and the before-and-after price math. The live tracker shows every model's current status.

Context too large

Token overflow (context_length_exceeded) and byte overflow (request_too_large) fail differently and need different fixes — counting tokens versus measuring payloads. When trimming isn't an option, the fix is a bigger window: the context-window comparison shows what each model's advertised window is really worth in practice.

Authentication and billing walls

Auth failures (401s) are the cheapest errors to prevent and the most embarrassing to debug at midnight. Gemini's FAILED_PRECONDITION is its own animal — a region and billing wall that hits the moment you deploy to a server outside free-tier coverage.

Server errors and overload

Anthropic's 529 and Gemini's 504 are the platform telling you about itself, not about your code. Retry politely, and if a provider's busy hours keep hurting you, the price-history record and rankings make the case for keeping a second provider wired as a fallback.