Errors & fixes
The errors developers actually hit while building against Claude, catalogued with what causes each one and the one-line fix — not a restatement of the API reference's parameter list. Each entry names the exact error shape you'd see in a response or a stack trace, so it's the kind of page worth bookmarking or linking from a Slack thread the moment you hit it, rather than one you'd read end to end.
Nearly every entry here traces back to one of two root causes: a parameter or behaviour that existed on an older model generation and was removed or changed on a newer one, or a limit — a token count, a breakpoint count, a lookback window — that a request quietly exceeded with no error at all, which is arguably worse than an error because nothing tells you it happened. The catalogue grows mostly from what readers report actually hitting in production, via contact, so if you hit something specific that isn't listed here yet, that report is the fastest way it gets added.
Where an entry involves a model-generation change specifically, it links to the fuller migration guide in /prompting/ — the error page is the fast fix; the migration guide is the full picture if you're moving a whole codebase rather than patching one call site.
budget_tokens is not supported on Opus 4.7 and later
temperature, top_p and top_k are not supported on this model
Assistant-turn prefill is not supported on this model
Disabling thinking above high effort on Opus 5
Fable 5 is unavailable to zero-data-retention organisations
The text-editor tool's type and name fields don't match
mcp_servers was set without mcp_toolset
Every tool was deferred to tool search, with none directly callable
The effort parameter errors on Haiku 4.5
xhigh effort errors on Opus 4.6 and earlier
cache_read_input_tokens is 0 with no error
speed: "fast" is not supported on Opus 4.7
Context window exceeded on a request you thought fit
max_tokens is too small once xhigh thinking is included
count_tokens gave a number that didn't match the bill
The model ID you're calling has been retired
The Batch API isn't available on your platform
More than 4 cache breakpoints in one request
A cache breakpoint missed because of the 20-block lookback limit
The model is deprecated but has no retirement date yet