Prime Radiant receipts · All receipts
Faster and cheaper on repeated questions
In plain language: when the same 300 questions are asked again, a stored card answers each one in about 4 to 6 milliseconds on an ordinary computer, with no AI model and no network. The most accurate flagship model took 35.71 seconds and billed $0.14778 per question for the same 300. If each of the 300 questions is asked 100 times, the cards cost about $99.30 in all against $4,433.46 for that flagship: 97.8 percent less.
The condition: repeated questions only. 300 questions, each asked 100 times. The first ask is a fresh call that mints a card; the next 99 replay it. On the first ask a card costs more than the flagship ($0.33 against $0.148). Nothing here measures new, never-seen questions.
Faster: replay time
| measurement | per question | source |
|---|---|---|
| Card replay on M0, recorded 2026-09-16 (300 questions, 300 of 300 matched the key, 0 model calls, 0 network attempts) | 4.38 ms | score_m0_vs_m2.json |
| Card replay on M2, the second machine, same run | 3.78 ms | score_m0_vs_m2.json |
| Re-run on M0 tonight, 2026-10-05, three times, same code and cards (same answers, same answers sha256 cf7ed835a5906dc8..., 0 model calls, 0 network attempts) | 5.43, 5.02 and 5.28 ms | recall_exec.cjs, re-run logs |
| Fresh call to the most accurate flagship (Anthropic), same 300 questions, 2026-09-16 | 35.71 s (10,712.1 s for 300) | scoreboard_300.json, re-scored tonight from the stored rows: same 35.71 s |
Tonight's replay was slower than the recorded 4.38 ms because the machine was busy with other work; the answers were identical. Either way a replay is thousands of times faster than a fresh call (35.71 s divided by 4.38 ms is about 8,150; divided by 5.43 ms it is about 6,580).
Cheaper: the arithmetic, line by line
- Flagship cost per question: Anthropic billed $44.3346 for the 300 questions, so $0.14778 per question. Source: scoreboard_300.json (re-scored tonight from the stored vendor rows: same $44.3346).
- Flagship, 300 questions asked 100 times: $44.3346 × 100 = $4,433.46 (the same as $0.14778 × 300 × 100, before rounding the per-question figure).
- Card mint cost: $0.33 per card, the first ask. This was measured on 100 of the 300 cards. The source records no mint cost for the first 150 cards; $0.33 stands in for every card in this sum. Source: D-BP223-BM0-X028-064-SCORE-300-HARNESS-BAND_REPORT.md.
- Card replay cost: $0.00001 per replay. This is a computed local cost (an estimate of the electricity for a few milliseconds of computing), not a meter reading. Same source.
- Cards, 300 questions asked 100 times: $0.33 × 300 + $0.00001 × 300 × 99 = $99.00 + $0.297 = $99.30.
- Difference: 1 − $99.30 / $4,433.46 = 97.8 percent less. Source of the arithmetic: D-BP223-BM0-X031I-X030-SESSION-READ-AND-HEADLINE-CLAIMS.md, section 2.2.
Limits
- Repeated questions only. The saving exists only when a question repeats exactly. A reworded question is a new question.
- Against Anthropic only. Anthropic was the most accurate and the most expensive per question. Against the other flagships the cards pay off later: after about 13 asks against OpenAI and about 28 against Google, and against Mistral (estimated price) the cards cost more. Source: section 2.2, break-even table.
- The cards were built for these 300 questions. 16 of the 300 were repaired with the answer key. This is exact recall of a stored method, not solving new questions. See the existing receipt The first 300 MMLU-Pro questions, 300 of 300 by code.
- The six flagships' figures are in the existing receipt Six flagships on the first 300. Re-scoring tonight reproduced every Anthropic figure. One other vendor's total rose by $0.1303 because two later DeepSeek calls were added to the stored rows after the first scoring; the six-vendor total is $65.7667 on tonight's re-score against $65.6364 recorded.
- One replay is compared with one fresh call. This is not a measure of a busy production system.
Source files (sha256)
| file | bytes | sha256 | what | check |
|---|---|---|---|---|
| score_m0_vs_m2.json | 1,329 | 5b39588137739016af2268c730fc90b2e2305fd8eb6345abca03261faa7d3d2c | replay time on M0 and M2, 300 questions | matches the receipt that lists it |
| answers_m0.json | 94,271 | 0635125076c9501a0267975cd883f80239e8cca3241523a638b868f507f2169a | per-question rows on M0 | matches the receipt that lists it |
| recall_exec.cjs | 16,424 | 5ec0548179375dedfdabc8b1e0371f078a719050ca8b523d2c5494cdf4543072 | the replay code (no model; refuses any network call) | matches the receipt that lists it |
| MANIFEST.json (card package) | 96,753 | abb77aa4bd667fa0681023eb6c8570d4e5809df17ab0c903ef96038f26097fcc | the 300 cards replayed | matches the receipt that lists it |
| scoreboard_300.json | 13,403 | 1c2f9dc191ea55159f43446b7ecbd316a48bc141514802633458c6431dc89c78 | six flagships on the same 300: billed dollars and seconds per question | matches the receipt that lists it |
| score_six_300.cjs | 10,591 | 7c027d9316090e66b8b775a4ba7898d99fa02ffb6f65a65d0c7105a6f9144050 | the scorer, reads stored vendor rows only | matches the receipt that lists it |
| D-BP223-BM0-X028-064-SCORE-300-HARNESS-BAND_REPORT.md | 8,929 | 8c15f4b6c905f523fc5b9b8f9894af059c37eb59d3e25e56a83452e370ffbada | the $0.33 mint cost (measured on 100 cards) and the replay estimate | not listed in an earlier receipt; hash recorded here |
| D-BP223-BM0-X031I-X030-SESSION-READ-AND-HEADLINE-CLAIMS.md | 25,463 | 4b887c2a70a0cf2fb10d47ea56489cb211db131c8a9a17980ea8a712b5736842 | the cost arithmetic, section 2.2 | not listed in an earlier receipt; hash recorded here |
Re-runs tonight: the replay code was run three times on M0 with the card package from the 2026-09-16 run, and the flagship scorer was run against the stored vendor rows with its output written to a separate folder. No vendor was called and nothing was spent. Built 2026-10-05T05:23:52Z.