Claude API cost: the tokenizer and the cache write

The tokenizer is a hidden price increase

Claude’s tokenizer emits roughly 43% more tokens than OpenAI’s o200k_base for the same English text. Providers meter their own tokenizer, so this is not an accounting curiosity — it is a real multiplier on every input line of your invoice.

Sonnet 5.5 sticker input     $2/1M
× tokenizer ratio            1.43
= effective input price      $2.86/1M of the same text

Any like-for-like comparison against another provider has to apply this, or it flatters Claude by about 43%. Our calculator applies it automatically; most published comparison tables do not.

The cache write costs 1.25× input

Anthropic bills first-time cache writes at $2.5/1M on Sonnet — 1.25× the standard input rate — where OpenAI and Google charge nothing extra to warm a cache. In exchange, cached reads come back at $0.1/1M, or 5% of input.

That is a genuine trade, not a penalty: pay a little more on the miss, get one of the steepest read discounts available on the hit. Which side you land on is entirely a function of your cache hit rate — high, steady traffic against a stable prefix wins, while spiky low-volume traffic pays the write premium repeatedly without earning it back. It’s worth modelling rather than guessing.

Sonnet or Haiku?

On the history-carrying chatbot above, Sonnet 5.5 runs $538/mo against Haiku 5.5 at $27/mo — Haiku is 95% cheaper for the same workload. Anthropic positions Haiku as near-frontier quality at budget pricing, which makes it an unusually strong default; the honest caveat is that “near-frontier” is a claim to verify on your task before committing at scale, not a spec to take on faith.