benchrShowing 15 of 15 errors
No errors match these filters. Clear one or more filters and try again.
The account has no usable quota. Retrying does not change the billing state.
rate_limit_exceededThe request or token rate exceeded the account's current limit.
context_length_exceededThe input and requested output together exceed the model's context window.
model_not_foundWrong ID, wrong endpoint, no access — or a model OpenAI retired.
invalid_api_key"Incorrect API key provided" — key wrong, revoked, or malformed.
overloaded_errorAnthropic is temporarily overloaded. Retry with jitter and a firm deadline.
rate_limit_errorTier ceiling — or acceleration limits if your usage ramped too sharply.
invalid_request_errorCheck unsupported sampling parameters, assistant prefills, and modified thinking blocks.
request_too_largeBody over the 32 MB Messages cap — rejected before Anthropic even sees it.
not_found_errorSince June 15, 2026 the top cause is a retired Claude model ID.
RESOURCE_EXHAUSTEDThe project exceeded a request, token, or quota limit for its current tier.
INVALID_ARGUMENTMalformed body — or a feature that doesn't exist on your API version.
NOT_FOUNDExpired file references — or a model from a line Google already shut down.
DEADLINE_EXCEEDEDThe request did not finish before the deadline. Reduce it, stream it, or adjust the timeout.
FAILED_PRECONDITIONFree tier unavailable in your region without billing enabled.
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: check the identifier and its lifecycle
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
For 401 errors, verify the key, project, organization, header, and any IP restrictions before changing application code. Gemini's FAILED_PRECONDITION can instead point to regional availability or a project that needs billing enabled.
Server errors and overload
Anthropic's 529 and Gemini's 504 usually need bounded retries, timeout handling, and monitoring rather than an immediate code rewrite. If availability is critical, use the price history and model reference to price and test a fallback before an incident.