One payment, one report. Five checks — caching, batching, right-sizing, retries, off-hours — each with its formula shown and its evidence attached. The demo dataset yields £1,114.78/mo.
If the report doesn't identify at least £99 of monthly savings, reply to the delivery email within 30 days for a full refund.
Headline sums HIGH and MED findings only. LOW sits under “possible, verify first”.
Stripe checkout. No account needed yet — the claim link in your receipt attaches the report to a new or existing account.
A read-only admin key (OpenAI, Anthropic, OpenRouter) or a CSV export from any provider console. Four minutes, no code.
Five checks over 30–90 days, each with the formula shown and the evidence it recomputes from. PDF and web, yours to keep.
The formula, plainly: saving = eligible_input_cost × cache_discount_rate × 0.6 adoption. The evidence table its arithmetic recomputes from. The confidence grade — HIGH means the money is there; MED states its adoption assumption; LOW is listed but never counted. And the fix, as numbered steps. The full worked example →