If your Anthropic bill jumped and you're trying to pin down why, it's usually one of a small set of explainable causes rather than a genuine mystery. Here's what we see most often behind an unexpected Claude API cost increase.
A cache write costs more than a plain input token — and if you never get a hit, you're paying extra, not saving
Anthropic's prompt caching charges a premium to write to the cache — roughly 1.25x a standard input token for a 5-minute cache, or 2x for a 1-hour cache — in exchange for a steep discount, often around 90%, on every cache read afterward. That trade only pays off if the same cached content actually gets reused. If your system prompt, tool definitions, or context change enough between calls that you rarely hit the cache, you're paying the write premium over and over without ever collecting the read discount, which can leave you worse off than not caching at all.
Extended thinking tokens land on the expensive side of the ledger
Like reasoning tokens on other providers, Claude's extended thinking tokens count toward output token usage — the pricier half of every request, typically several times the cost per token of input. Turning on deeper extended thinking, or letting a model reach for it more often on harder prompts, raises your effective cost per response even if the final answer you see looks about the same length as before.
You crossed into a different usage tier — or hit its spend cap
Anthropic moves organizations through usage tiers as usage history accumulates, and each tier carries its own monthly spend ceiling. Crossing into a higher tier doesn't change your per-token price, but it does raise how much you're able to spend before being capped — and on a lower tier, hitting that spend ceiling doesn't just slow requests down, it can pause API access outright until the next billing cycle. Either direction can look like a billing surprise if you're not watching which tier you're on.
Retried requests after a 429 add up fast
A rate-limited request returns an HTTP 429 rate_limit_error with a retry-after header telling you how long to wait. A retry loop that ignores that header, or that doesn't back off, can resend the same logical call multiple times — and each attempt is billed independently.
Anthropic adjusted pricing on a model or mode
Anthropic has both cut and held pricing steady across model updates — for example, a faster processing mode on a newer Claude model launching at a fraction of a previous mode's per-token rate. These changes don't always come with the same fanfare as a new model release. Check the Anthropic calculator above for the latest verified rate and price history, and add a free watch so the next change reaches you before your invoice does.