D2 — Missing cache

The problem

Both major providers discount cached prompt prefixes steeply — Anthropic cache reads cost a tenth of the input rate. If your traffic re-sends the same system prompt or context block at the full input rate, you are paying the same bill over and over for tokens the provider would happily serve from cache.

Detection

The detector groups calls by identical prompt prefix (SHA-256 over the first 4,096 prompt characters when your export carries prefix_hash; a conservative heuristic otherwise) and finds groups that never used cache despite repeating within the provider's cache lifetime.

Savings estimate

The estimate is a full cache simulation, not a flat discount:

cacheable    = hash-verified: min(prompt_tokens, 1024-token cap)
               heuristic:     0.8 × group-minimum prompt
est_writes   = number of cache-lifetime windows the group spans
               (one write per window; TTL is provider-family specific)
gross        = reads × cacheable × (input_rate − cache_read_rate) / 1e6
penalty      = est_writes × cacheable × (cache_write_rate − input_rate) / 1e6
savings      = gross − penalty        (0.7 haircut if windows can't be estimated)

Cache lifetimes differ per provider family — Anthropic's 5-minute TTL means more re-writes than the GPT-5.6 family's 30-minute minimum, and the write premium differs too (Anthropic writes at 1.25× input; the 5.6 family likewise; older OpenAI families have no premium). The simulation uses the right window and premium per family.

Worked example (golden fixture)

30 uncached claude-sonnet-5 calls, identical 5,400-character prefix, 2,000 prompt tokens, spaced 130 seconds apart (intro rates: input $2, read $0.20, write $2.50; TTL 300s):

windows   = 13  →  est_writes 13, cached reads 17
cacheable = min(2000, 4096/4) = 1024 tokens (hash-verified cap)
gross     = 17 × 1024 × (2.00 − 0.20) / 1e6 = $0.0313344
penalty   = 13 × 1024 × (2.50 − 2.00) / 1e6 = $0.0066560
observed  = $0.0246784   →  monthly (×10) = $0.246784

Note what the conservatism does: a naive "cache everything" estimate would claim the full read discount on all 30 calls; the simulation pays for 13 cache writes first and caps cacheable tokens at the hash-verified span.

Known limitations

OpenAI usage exports do not carry per-call cache-write counts (TRACKED GAP): for OpenAI traffic the write penalty is estimated from TTL windows, never observed. Anthropic exports carry both cache fields natively. If a prefix repeats but crosses cache-lifetime boundaries slowly, real-world savings will be lower than a same-minute burst — the window simulation models exactly this.