API prices

<100.00% of input tokens billed as cache hits, <100.00% as cache writes, <100.00% at the plain input rate.98% of input tokens billed as cache hits, 2% as cache writes, 0% at the plain input rate. Measured, not assumed: the token-weighted mean of 64,680 recorded coding-agent requests. Almost every billed input token is a cache read, because a conversation re-reads its whole prefix every turn while writing only that turn’s delta — reads grow with the square of the turn count and writes only linearly. Output is 0.4% of input, so this is a prefill-bound workload, not a decode-bound one.84% of input tokens billed as cache hits, 15% as cache writes, 1% at the plain input rate. A short exchange over a stable system prompt. At around a dozen turns the arithmetic gives 15% write and 85% read, which matches the shortest real sessions in the recorded corpus. Caching may not engage at all on models with a 4,096-token minimum cacheable prefix.76% of input tokens billed as cache hits, 20% as cache writes, 4% at the plain input rate. One document written to cache once, then re-read for each follow-up question. Assumes five questions per document — an assumption, not a measurement, and the number here most worth changing. At a single question caching cannot pay for itself and the honest mix is no caching at all.84% of input tokens billed as cache hits, 15% as cache writes, 1% at the plain input rate. The same shape as long-document reading at the top of a million-token window, where the write is worth amortising over more questions — twelve, on the same arithmetic. Also the case where KV cache, not weights, decides how much hardware a self-hosted deployment needs.60% of input tokens billed as cache hits, 5% as cache writes, 35% at the plain input rate. A short brief against a shared style guide, producing a long artefact — the mirror image of agentic coding and the only preset here where output dominates the bill. The input mix barely moves the result. 98% of input tokens billed as cache hits, 2% as cache writes, 0% at the plain input rate. Measured, not assumed: the token-weighted mean of 64,680 recorded coding-agent requests. Almost every billed input token is a cache read, because a conversation re-reads its whole prefix every turn while writing only that turn’s delta — reads grow with the square of the turn count and writes only linearly. Output is 0.4% of input, so this is a prefill-bound workload, not a decode-bound one.