A call's estimated cost prices tokens written to a cache at the cache-read rate #50

Priority

high

Components

• app/nullius (backend)
• app/web (routes)
• templates and static files
• docs and the site
• catalogue (benchmarks, agents, capabilities)
• the standard (ports, kits, relay, runs)
• CI, workflows and skills

Evidence

app/nullius/standard/relay.py, reading a reply's usage, adds the tokens written to a cache to those read from it:

cache = (u.get("cache_read_input_tokens") or 0) + (u.get("cache_creation_input_tokens") or 0)

and _cost prices all of them at the cache-read rate:

fresh * p["input"] + usage["cached"] * p.get("cached", p["input"]) + usage["output"] * p["output"]

A price row carries cache_write (app/nullius/prices.py, from the list's cache_creation_input_token_cost), which is not used. On Anthropic's prices a cache write costs 1.25 times the input rate and a cache read 0.1 times: tokens written to the cache are counted at a twelfth of their bill. Found by a review of the spending cap, not by a run.

Reproduction

1. A standard run on a provider whose replies carry cache_creation_input_tokens and no cost of their own, with a model whose price row has cached and cache_write.
2. Compare a call's cost_usd in its calls.jsonl with the provider's bill for it.

Expected

Tokens written to a cache are priced at cache_write, those read from it at cached.

Acceptance

• A reply's cache-write tokens are priced at the row's cache_write rate (the input rate when the row has none), apart from its cache-read tokens.
• A test with a reply carrying both kinds shows the estimate.

Dependencies and related

A run's spending cap (#14, #46) settles a call at this estimate when the reply carries no cost of the provider's own, so on such a provider it can let a run spend past its cap. Placeholder priority: the lowest defensible, it changes a number shown, not what runs.