# The cost of intelligence, as a utility: the receipts table

*Cold scroll: what this file is.* Every figure on the page [*The cost of intelligence, as a utility*](/the-cost-of-intelligence/), with its value, the conditions that matter (gates, flags, postures, in the value cell) and the kit file and line it was typed from. It ships in the data kit beside the page; the page links it from "How to check our work". Paths are inside the public kit, whose root is `/the-cost-of-intelligence/data/` (its `index.json` lists every file); line numbers are each file's in the bytes the kit's `index.json` pins for it by sha256. Rows of the series are cited by model, card and set, never by row id: the public kit withholds the hashed row ids. Ids of the form `R-`, `A-`, `O-` and `F-` name the numbered rules of the pre-registration, its amendments and the metrics file, `P-` ids the house price rules the pre-registration adopts, and `CB-` labels the pre-registration's later cells; the kit carries screened copies of those files.

## Reading 0 and the first reading-0.1 cells (the page's figures)

| Figure | Value | file:line |
|---|---|---|
| The six-way set | 108 items, six guests, public 2026-09-19 / 2026-09-23T13:16Z; S0, frozen | `SERIES-SETS.md:7`; `PREREG-S0.md:31` |
| The sets' public dates | Doorman never published; field exam's source items 2026-08-29; docent gate 2026-09-23 (`SERIES-SETS.md:9` disagrees on the doorman; that file was not corrected at the kit commit of 2026-10-01) | `PREREG-S0.md:31–34` |
| Not-clean rows | 14 unverified (no receipted weights date), 1 possible; no ratio uses them | `series.public.csv`, column `contamination` |
| Reading 0's window | 2026-09-23T15:40:04Z (the shelf's first row) → 2026-09-25T12:00Z (the registered ceiling); the registered opening stamp is 15:43:21Z, the smoke's first scored row | `cloud-reading-0/CLOUD-READING-0.md:5`; `PREREG-S0.md:426` |
| Reading 0 row counts | 339 rows; 247 headline (48 provisional); A-05 | `PREREG-S0.md:514` |
| The anchor, six-way | 88/108 (81.5 %, Wilson 73.1–87.7 %); $0.9553; $0.01086 per correct; $30/$60 per M; temperature 0, seed 20260921 | `openai-reading-0/OPENAI-READING-0.md:203, :107`; `series.public.csv` row: (gpt-4-0613`, OpenAI's API, six-way) |
| The bridge id | `gpt-4o-mini-2024-07-18` 73/108; callable at close | `OPENAI-READING-0.md:208` |
| The 3.5 line | 34, 24 of 108 (the `-instruct` row, 15, was not in the series at the kit commit of 2026-10-01); `-1106` 2023-11-06, `-0125` 2024-01-25 | `OPENAI-READING-0.md:204–205, :108–109`; a workshop ruling of 2026-09-25 |
| Top OpenAI row | `gpt-5.6-sol` 103/108; effort none, temperature 0 | `OPENAI-READING-0.md:228`; row (gpt-5.6-sol`, OpenAI's API, six-way) |
| The default-effort re-reads | `gpt-5.5` 103 at $0.4691 vs 102 at $0.1609 (effort none); server-echoed default = medium | `OPENAI-READING-0.md:225–226` |
| Default-temperature OpenAI ids | 7 ids (9 rows) at "default (1)": the reasoning ids refuse a fixed temperature | `OPENAI-READING-0.md:211, :215–219, :230–231` |
| OpenAI spend, reading 0 | $4.4277 at list price | `OPENAI-READING-0.md:382` |
| Cheapest OpenAI at or above the anchor | `gpt-6-luna` 96/108, $0.0032, $0.0000333 per correct, ≈ 326× cheaper per correct than the anchor; 274 input tokens per item; API created 2026-09-14, changelog 2026-09-22 | `OPENAI-READING-0.md:47–50, :136, :232, :274`; row (gpt-6-luna`, OpenAI's API, six-way); `OPENAI-READING-0.md:245` (the anchor: 30,544 in over 108) |
| First local clearer, reference card | `qwen3.5:27b` 97/108; lower bound +0.0177; 300 W, think off, 8 tokens, Q4_K_M, 4,096 ctx; both halves; 37 read | `charts/data/c3-ladder.csv:8`; `c3-tests.csv:93`; row (qwen3.5:27b-q4_K_M`, the reference card, six-way, reasoning off) |
| First local clearer, laptop | `qwen3.5:27b` 96/108; lower bound +0.0096; both halves; 32 read | `c3-ladder.csv:9`; `c3-tests.csv:125`; row (qwen3.5:27b-q4_K_M`, the laptop card, six-way, reasoning off) |
| The star alternative | `qwen3.6:27b` 96 / 95; lower bounds +0.0294 / +0.0171; weights 2026-04-21; 2026H1 ★, chosen before the v2 re-score | `c3-tests.csv:97, :129`; rows (qwen3.6:27b-q4_K_M`, the reference card, six-way, reasoning off), (`qwen3.6:27b-q4_K_M`, the laptop card, six-way, reasoning off); `PREREG-S0.md:510` (A-11) |
| The "scored above, does not clear" rows | `qwen2.5:32b` 89 / 90, lower bounds −0.0648 / −0.0734; `gemma3:27b` 89 / 89, −0.0901 / −0.0909; δ = 5 points at n ≥ 100, 10 at n < 100 | `c3-tests.csv:79, :112, :83, :115`; `PREREG-S0.md:305` |
| When the bar was fixed | R-15 ~17:3xZ 2026-09-23; committed 18:34:55Z (`2cd9bb79`); arms' `started`, reference card: `qwen2.5:32b` 17:46:14Z, local anchor 17:52:46Z (whole set only), `gemma3:27b` 18:09:59Z, `qwen3.5:27b` 20:16:45Z, `gemma4:31b` 20:26:22Z, `qwen3.6:27b` 20:38:01Z, `qwen3.8:27b` 23:19:18Z; laptop `qwen3.5:27b` 21:18:58Z, `gemma4:31b` 21:26:12Z, `qwen3.6:27b` 21:33:16Z; A-00 2026-09-24T13:18:41Z; rented six-way rows: 56 of 61 started before the commit (OpenAI 29 of 29, 16:00:42Z–18:03:37Z; cloud 20 of 20; Claude 7 of 12), 30 of the 34 rented both-halves clearers among them | `RULINGS-2026-09-23.md:229–252`; the arms' `arm.json`, field `started`: on the reference card [`b-qwen2.5-32b`](/the-cost-of-intelligence/data/rtx-3090-reading-0/rows/b-qwen2.5-32b/arm.json), [`k0a-gemma4-anchor-cd`](/the-cost-of-intelligence/data/rtx-3090-reading-0/rows/k0a-gemma4-anchor-cd/arm.json), [`b-gemma3-27b`](/the-cost-of-intelligence/data/rtx-3090-reading-0/rows/b-gemma3-27b/arm.json), [`b-qwen3.5-27b`](/the-cost-of-intelligence/data/rtx-3090-reading-0/rows/b-qwen3.5-27b/arm.json), [`b-gemma4-31b`](/the-cost-of-intelligence/data/rtx-3090-reading-0/rows/b-gemma4-31b/arm.json), [`b-qwen3.6-27b`](/the-cost-of-intelligence/data/rtx-3090-reading-0/rows/b-qwen3.6-27b/arm.json), [`b-qwen3.8-27b`](/the-cost-of-intelligence/data/rtx-3090-reading-0/rows/b-qwen3.8-27b/arm.json); on the laptop [`b-qwen3.5-27b`](/the-cost-of-intelligence/data/laptop-reading-0/rows/b-qwen3.5-27b/arm.json), [`b-gemma4-31b`](/the-cost-of-intelligence/data/laptop-reading-0/rows/b-gemma4-31b/arm.json), [`b-qwen3.6-27b`](/the-cost-of-intelligence/data/laptop-reading-0/rows/b-qwen3.6-27b/arm.json); `PREREG-S0.md:498`; `openai-reading-0/OPENAI-READING-0.md:29–30`; each rented row's `report.json` `started` (`openai-reading-0/rows/`, `cloud-reading-0/rows/`, `anthropic-reading-0/rows/`) |
| The resampling correction | 14 verdicts moved (11 against the anchor, 3 against the local anchor) when the 16-dinner unit replaced single lines, 2026-09-27; laptop `qwen2.5:32b` lower bound −0.0463 (lines, clears) → −0.0734 (dinners, fails); both halves read for every row, asked of the cheapest clearer only | `F0-NOTES.md` §1, §3, §4; `PREREG-S0.md:311`; the correction ruling of 2026-09-27 |
| The chance bar | 27 of 37; 24 of 32; first `qwen:32b-chat-v1.5` 2024-04-03, 62 / 61 of 108; the shelf 19 of 20, first `gpt-oss:120b` 83/108; floor 24.0 %; the ratio's correction is 1/6 | `c3-ladder.csv:2–3, :7`; `PREREG-S0.md:271, :273` |
| The shelf at the anchor's bar | 20 read, 11 clear; first `mistral-large-3:675b` 2025-12-02, 91/108; plan-included | `c3-ladder.csv:13` |
| The 96 GB card at reading 0 | 0 of 5 clear; best `gemma4:26b` 86/108; 3 ratio-eligible; 420 W, persistence off, beside the services | `c3-ladder.csv:10`; `charts/data/c4-numerator.csv`; row (gemma4:26b-a4b-it-q4_K_M`, the 96 GB card, six-way, reasoning off) (`power_posture`) |
| The staircase, nominal | 298.65-fold, 3.2548 y, 5.761×/y, 35.45 %/quarter, 9 steps; launch price, $ per item | `charts/data/c2-staircase-rates.csv:2`; `F1-NOTES.md` §5 |
| The staircase, real | 328.24-fold, 5.930×/y, 35.92 %/quarter, 10 steps (`gpt-5.1-2025-11-13` added at 1.00); `slower_than_all_ages` True; CPI-U, base 2026-08, the last step at the base month | `c2-staircase-rates.csv:3`; `c2-staircase-steps.csv:18, :20` |
| The last-step sensitivity | Factor 10.18 (two-fifths of the fall in log terms); without it 29.34-fold, 3.053×/y (32.12-fold, 3.146×/y real); today's prices 146.71-fold, last step 2.04, `gpt-5.6-luna` cut ×5.00 | `c2-staircase-steps.csv:10, :28–29`; `c2-staircase-rates.csv:2–4` |
| Epoch's rate | 47.0 %/quarter = 12.67×/y, real dollars, 2023–26 all ages; pinned `Figure 1.csv`, CC BY 4.0 | `charts/data/c2-epoch-rates.csv:2` |
| Epoch's age-matched rate | Not registered; "≤ 4.7×/year" is plan prose | `c2-epoch-rates.csv:3`; `PLAN-v2.md` §1; `recon/economics.md:68` |
| Epoch's citation | Authors, date, URL, licence, the requested form | `recon/economics.md:54–57`; `PLAN-v2.md:328` |
| Survivorship | `gpt-4-0613`, `gpt-4-turbo-2024-04-09`, `gpt-4o-2024-05-13`, `gpt-3.5-turbo-0125`, `o3-mini`, `o4-mini`, `gpt-4.1-nano` retire 2026-10-23; `gpt-5`, `o3`, `gpt-5-mini`, `gpt-5-nano` 2026-12-11; `gpt-3.5-turbo-1106` 2026-09-28; a stopgap list read 2026-09-23T15:40:18Z | `charts/data/c2-survivorship.csv:3–19` |
| The ratio, reference card, fixed | 1.1143 (1.0160–1.2463); L 0.8889, F 0.8148; with its spread (the operators' wording ruling of 2026-09-27); seed rows §4, §8 | `F1-NOTES.md` §3; `charts/data/c4-ratios.csv:3`; `charts/data/ratio-seed-sensitivity.csv` |
| The ratio, reference card, moving | 0.9176 (0.8341–1.0127), k = 14 (7 Claude rows); provisional | `F1-NOTES.md` §3; `c4-ratios.csv:2`; `charts/data/c4-f-closed.csv` |
| The ratio, laptop | Fixed 1.0571 (0.9873–1.2097); moving 0.8706 (0.8153–1.0000); the pick `gemma4:31b`; both provisional, at the boundary | `F1-NOTES.md` §3; `c4-ratios.csv:6–7` |
| The ratio, laptop, home-team split | Fixed-split 1.1094 (0.9176–1.2264), not provisional; moving-split 1.0718 (0.8806–1.1765), provisional; 57 items; the split rows re-pick on their items | `F1-NOTES.md` §3; `c4-ratios.csv:10–11`; `PREREG-S0.md:286` |
| The ratio, laptop, open-cloud and temperature-0 | 0.9867 (0.9054–1.1018); 0.8706 (0.8155–0.9883), position below; both provisional (the pick) | `F1-NOTES.md` §3; `c4-ratios.csv:8–9` |
| The ratio, temperature-0, reference card | 0.9176 (0.8345–1.0000); greedy frontier rows only (k = 6) | `F1-NOTES.md` §3; `c4-ratios.csv:5` |
| The ratio, 96 GB card | Fixed 0.9143; moving 0.7529; roster coverage, not the card (k = 3) | `F1-NOTES.md` §3 |
| The open-cloud denominator | `kimi-k3` 2,812 B, `deepseek-v4-pro` 1,650.5 B, `kimi-k2.7-code` 1,042 B; reference card 1.0400 | `F1-NOTES.md` §3 |
| Local energy rows | 183.4 / 231.0 / 233.6 J per correct on the reference card; 28.6 / 58.0 / 61.2 on the laptop; compute phase 26.3 / 71.4 / 75.1 and 7.8 / 27.7 / 30.0 (`j_per_correct_board_compute`); mean W as tabled; BOARD, strict | `series.public.csv` rows: (gemma4:26b-a4b-it-q4_K_M`, the reference card, six-way, reasoning off), (`qwen3.5:27b-q4_K_M`, the reference card, six-way, reasoning off), (`qwen3.6:27b-q4_K_M`, the reference card, six-way, reasoning off), (`gemma4:26b-a4b-it-q4_K_M`, the laptop card, six-way, reasoning off), (`qwen3.5:27b-q4_K_M`, the laptop card, six-way, reasoning off), (`qwen3.6:27b-q4_K_M`, the laptop card, six-way, reasoning off) |
| The 96 GB card's anchor energy | 157.4 gross / 74.3 net / 12.7 compute-phase J per correct, beside the services | the 96 GB card's beside-the-services tables, [`rtx-pro-6000-reading-0/TABLES-beside.md`](/the-cost-of-intelligence/data/rtx-pro-6000-reading-0/TABLES-beside.md), line 7 |
| Laptop vs reference card energy | Median 0.27× (0.06–0.62) whole window; 0.42× (0.11–0.74) compute phase, the fair pair; 0.50× per generated token (33 models); BOARD vs BOARD, 67 shared cells | `laptop-reading-0/LAPTOP-READING-0.md:21` |
| Same file, different answers across cards | Local anchor 88 vs 84; 102 of 108 agree, bridge line 105; 6 of 6 clean disagreements; 38 of 74 cells carried | `LAPTOP-READING-0.md:19–20` |
| The calibration run's joules | 176.3 vs 194.3 J per correct on byte-identical answers; 300 W, one card, vs 250 W with a second card on the bus; `num_ctx` 131,072; "a posture and box difference"; resident idle 106.4 W | the reference card's reading-0 report, [`rtx-3090-reading-0/RTX-3090-READING-0.md`](/the-cost-of-intelligence/data/rtx-3090-reading-0/RTX-3090-READING-0.md), lines 63 (`num_ctx` 131,072), 68, 78, 80, 85 and 95–98 |
| The Claude line | In release order (the page prints eight of the twelve ids; this row carries all twelve): `claude-sonnet-4-5` 2025-09-29, 97 of 108; `claude-haiku-4-5` 2025-10-15, 81; `claude-opus-4-5` 2025-11-24, 100; `claude-opus-4-6` 2026-02-05, 100; `claude-sonnet-4-6` 2026-02-17, 98; `claude-opus-4-7` 2026-04-16, 97; `claude-opus-4-8` 2026-05-28, 101; `claude-fable-5` 2026-06-09, 53; `claude-sonnet-5` 2026-06-30, 102; `claude-opus-5` 2026-07-24, 98; `claude-fable-5-1` 2026-09-01, 95; `claude-opus-5-5` 2026-09-22, 104; CLI default effort; one scored pass (haiku and opus-4-8 two); the account's training setting unreadable | `series.public.csv` rows: (claude-sonnet-4-5`, the command-line tool, six-way, default effort), (`claude-haiku-4-5`, the command-line tool, six-way, default effort), (`claude-opus-4-5`, the command-line tool, six-way, default effort), (`claude-opus-4-6`, the command-line tool, six-way, default effort), (`claude-sonnet-4-6`, the command-line tool, six-way, default effort), (`claude-opus-4-7`, the command-line tool, six-way, default effort), (`claude-opus-4-8`, the command-line tool, six-way, default effort), (`claude-sonnet-5`, the command-line tool, six-way, default effort), (`claude-opus-5`, the command-line tool, six-way, default effort), (`claude-fable-5-1`, the command-line tool, six-way, default effort), (`claude-opus-5-5`, the command-line tool, six-way, default effort), (`claude-fable-5`, the command-line tool, six-way, default effort) |
| Claude API-equivalent $ per correct | sonnet-4-5 $0.00909; opus-5-5 $0.00313; sonnet-5 low $0.000717; derived from the published list, scaffold removed; not money paid | The same rows and (`claude-sonnet-5`, the command-line tool, six-way, low effort), `usd_per_correct_today` |
| Fable 5's refusals | 53 of 108, all "cyber"; 53 of 55 answered right; $1.8923 billed for refused calls of $2.9608 (the tool's list-price estimate); R-21 fallback off; scored wrong | `anthropic-reading-0/ANTHROPIC-READING-0.md:136, :142`; row (claude-fable-5`, the command-line tool, six-way, default effort) |
| Claude sampling noise | Haiku 79/108 agree (73.1 %); opus-4-8 108/108; the tool exposes no temperature | `ANTHROPIC-READING-0.md:330–331, :326` |
| Claude spend | $21.1936, the tool's list-price estimate; never enters a ratio | `ANTHROPIC-READING-0.md:5, :470`; `PREREG-S0.md:362` |
| The cloud shelf | 8,336 calls, 5 failed, 0 × 429; calibration 68.5 % = 68.5 %, agreement 73.1 % vs floor 72.2 %; plan-included | `cloud-reading-0/CLOUD-READING-0.md:367, :325–326` |
| The think-only star | `deepseek-r1:32b` on 1/108 strict, 83 lenient, 696,094 J per correct, 2,048 budget at 8,192 ctx, capability row (provisional); off 0/108, the protocol's verdict, not a headline row; no amendment rules "not a reading" | Rows: (`deepseek-r1:32b-qwen-distill-q4_K_M`, the reference card, six-way, reasoning on), (`deepseek-r1:32b-qwen-distill-q4_K_M`, the reference card, six-way, reasoning off); `A05-NOTES.md:157` |
| Reasoning budgets | Fast pairs 4,096 at 8,192 ctx; five dense rows 2,048; three off rows at 2,048 ctx; 179 of 183 budget-stopped replies empty | the reference card's reading-0 report, [`rtx-3090-reading-0/RTX-3090-READING-0.md`](/the-cost-of-intelligence/data/rtx-3090-reading-0/RTX-3090-READING-0.md), section 4, lines 257–266; `series.public.csv` column `num_ctx` |
| The buyer's inputs | Capital: the whole rig's replacement price read 2026-09-13 (*The cost to purchase and run a very capable home AI rig*; provisional, its card line struck 2026-09-18 (O-12); the price page's correction of 2026-09-18 re-reads the same listing); electricity: the U.S. average residential price for June 2026, EIA *Electric Power Monthly* table 5.6.A, read 2026-09-23 (*What it costs to run a high-end AI rig at home* uses July's, published the next day); a three-year straight-line life; the laptop's fuel the board's joules ÷ 0.9, assumed, its box idle unmeasured (series `cost_basis`; O-13); six hours a day (a quarter of the hours) or around the clock; no price prints here, by this page's rule: both figures are public on the pages named, which this page links | `PREREG-S0.md:353`; `METRICS.md:157–158, :259, :269, :446–449 (O-12), :450`; the rig's price: *The cost to purchase and run a very capable home AI rig* (/home-inference-and-diffusion-rig-cost/), the public price page |
| Break-even cost ratios | Reference card 5.705 / 1.676; laptop 2.633 / 0.724 (`qwen3.8:27b`, at the boundary, fails one half: provisional by `PREREG-S0.md:311`; the both-halves clearer `qwen3.5:27b` 2.719 / 0.744, `ratio-cost-ratio.csv:11`); at U25: n5 3.809, i5 6.187, both 4.249, commercial 5.487, industrial 5.223; **provisional** (the 2026-09-13 whole-rig price; its card line struck 2026-09-18; O-12); verdict withheld | `F1-NOTES.md` §6; `charts/data/ratio-cost-ratio.csv` row `rtx-3090,B-anchor`; `ratio-ustar.csv` |
| Break-even counts at the chance bar | `u*` ≤ 1 on 4 of 27 (reference card), 20 of 24 (laptop); `C*` ≤ 0 on 7 (reference card); archetype A | `F1-NOTES.md` §6 |
| The registered prediction | Cloud falls faster at the anchor's bar by 1 to 1.6 a year; forward readings only; the band's numbers by amendment A-06, and while they are unregistered (as at reading 0.1's close, 2026-10-14T00:00Z) the rule prints "band not yet registered"; no ×/year under 12 months; reading 1 2026-12-23 | `PREREG-S0.md:366`; `PLAN-v2.md` §1 |
| Reading 0.1's definition | Not on reading 0's list, plus listed cells that slipped past the close; reading 0's rows frozen in it | `PREREG-S0.md:426, :518 (A-13), :520 (A-14)` |
| The first big-model window (0.1) | `falcon:180b` 0/108 strict, 19 lenient (100 late letters, 8 word letters); 0/48 on the doorman; 2026-09-24, arms 06:35:55Z–06:37:43Z (smoke) and 06:38:16Z–06:45:56Z (falcon) (`arm.json` `started`, `ended`); two cards; contamination unverified | Rows: (`falcon:180b`, the 96 GB card with a second card, six-way, reasoning off), (`falcon:180b`, the 96 GB card with a second card, doorman's set, reasoning off); both arms' `arm.json`: [`s-mistral-v0.1-spread-cd`](/the-cost-of-intelligence/data/rtx-pro-6000-reading-0/rows/s-mistral-v0.1-spread-cd/arm.json) and [`f-falcon-180b-cd`](/the-cost-of-intelligence/data/rtx-pro-6000-reading-0/rows/f-falcon-180b-cd/arm.json); the window ruling of 2026-09-24 |
| The anchor's repeat and other sets (0.1) | 106/108; 88 = 88; 0.0 [−0.0270, +0.0275]; ≥ 105 HELD; 2 disagreements < 3; j 91/119 (76.5 %), "under 50 %" FAILED, the six j ids 72.3–81.5 %; f40 40/40; `o1` six-way 97/108; task_a 39/63, 36/36 at 8k, 24 pages over the window; $6.467161 and $3.3678 at list price | `openai-reading-0/CB1-PART1-REPORT.md:1, :32–33 (the window), :158, :170, :173, :266–274`; `openai-reading-0/receipts/cb1-part2/FIRST-MEASURE.md:17 (15:10:04Z), :20–21 (15:14:27Z), :23–28` |
| Claude low effort (0.1) | 100.00, 98.33, 79.67 (means of 3); fable-5-1 −15.33 [−23.66, −7.40] answers; lenient 101.0; refused 0; CB-2-R1 (a): the default row stays (the ruling of 2026-09-27) | `anthropic-reading-0.1/ANTHROPIC-READING-0.1.md:92–94` |
| Claude on the doorman (0.1) | fable-5-1 refused 6 of 6 sent; opus-5-5 refused 6 of 48; the set is written to provoke refusals | Rows: (`claude-fable-5-1`, the command-line tool, doorman's set, default effort), (`claude-opus-5-5`, the command-line tool, doorman's set, default effort) |
| CB-10 (0.1) | 60 cells; 54 collected; 6 refused; $5.45283; sealed, not judged; window 08:49:49Z to 08:56:25Z | `quality-q3-anchor/JUDGING-STAGED.md:1–7` (the window at `:7`); the dollars at `openai-reading-0/receipts/cb1-part2/FIRST-MEASURE.md:30, :117` |
| The seconds per answer behind the owned dollars | `t_busy` = the task window ÷ 108, one question at a time: `qwen3.8:27b` 1.028 s on the reference card (load step median 507 ms) and 0.547 s on the laptop (157 ms); the local anchor 0.963 s (768 ms) and 0.399 s (242 ms); not defined in the pre-registration (O-11); no compute-paced reading had been computed at the kit commit of 2026-10-01 | `series.public.csv` rows: (qwen3.8:27b-q4_K_M`, the reference card, six-way, reasoning off), (`qwen3.8:27b-q4_K_M`, the laptop card, six-way, reasoning off), (`gemma4:26b-a4b-it-q4_K_M`, the reference card, six-way, reasoning off), (`gemma4:26b-a4b-it-q4_K_M`, the laptop card, six-way, reasoning off), columns `cost_basis` (`t_busy`) and `load_ms_median`; `METRICS.md:137–138, :443–445`; `PLAN-v2.md:90` |
| The loading step inside the window | Reference card, local anchor: task window 104.0 s at a mean 155.1 W; the 108 calls' `x_load_ms` sum to 82.4 s (median 768 ms) and `x_prompt_eval_ms + x_eval_ms` to 14.9 s (0.14 s a question); laptop: 43.0 s, 26.3 s (median 242 ms), 11.8 s; "a property of this server and must never be carried to another one" | the two anchor arms' `.c.report.json` (`energy.seconds`, `energy.mean_watts`) and `.c.jsonl` rows: the reference card's [`k0a-gemma4-anchor-cd.c.report.json`](/the-cost-of-intelligence/data/rtx-3090-reading-0/rows/k0a-gemma4-anchor-cd/k0a-gemma4-anchor-cd.c.report.json) and [`k0a-gemma4-anchor-cd.c.jsonl`](/the-cost-of-intelligence/data/rtx-3090-reading-0/rows/k0a-gemma4-anchor-cd/k0a-gemma4-anchor-cd.c.jsonl); the laptop's [`k0a-gemma4-anchor-cd.c.report.json`](/the-cost-of-intelligence/data/laptop-reading-0/rows/k0a-gemma4-anchor-cd/k0a-gemma4-anchor-cd.c.report.json) and [`k0a-gemma4-anchor-cd.c.jsonl`](/the-cost-of-intelligence/data/laptop-reading-0/rows/k0a-gemma4-anchor-cd/k0a-gemma4-anchor-cd.c.jsonl); the reference card's reading-0 report, [`rtx-3090-reading-0/RTX-3090-READING-0.md`](/the-cost-of-intelligence/data/rtx-3090-reading-0/RTX-3090-READING-0.md), lines 107–111; `METRICS.md:68` (apportioned; O-4 at `:423–425`) |
| The cheapest clearer's own row | `qwen3.8:27b` weights 2026-08-05, clean; 95 of 108 on the reference card (both halves); 93 on the laptop (even half's lower bound −0.1304 against −0.10: fails; at the boundary) | `c3-tests.csv:99, :131`; `c3-ladder.csv:8–9`; rows (qwen3.8:27b-q4_K_M`, the reference card, six-way, reasoning off), (`qwen3.8:27b-q4_K_M`, the laptop card, six-way, reasoning off) |
| The both-halves laptop figure | `qwen3.5:27b` 2.719 at six hours a day and 0.744 around the clock against `gpt-6-luna` at today's price (the model's own cost ratio; the row is filed under the margin bar); clears the anchor's bar on both halves | `charts/data/ratio-cost-ratio.csv:11`; `c3-tests.csv:125`; rows (qwen3.5:27b-q4_K_M`, the laptop card, six-way, reasoning off), (`gpt-6-luna`, OpenAI's API, six-way), columns `usd_per_correct_A_U25`, `usd_per_correct_A_U100`, `usd_per_correct_today` |
| The tool's refusal fallback | By default the command-line tool re-sent a refused item to another Claude model, naming it in its own usage record and flagging the re-send (`model_refusal_fallback`); the served-model check stopped the model on the first day; re-sent items out of every rate; proof with the fallback off 2026-09-23T21:09:31Z | `anthropic-reading-0/ANTHROPIC-READING-0.md:14, :16, :127, :138, :268, :477`; `anthropic-reading-0/receipts/refusal-proof.json` |
| The private set's egress | 48 lines to the shelf; directly to six Claude ids, first call 2026-09-23T15:48:50Z, last 17:39:06Z, the hold at about 17:38Z; 12 re-sent lines whose usage record names a seventh id, `claude-opus-5`; none to OpenAI | `anthropic-reading-0/receipts/doorman-egress.json` (`direct`, `fallback_resends`, `why`) |
| The families that wrote the lines | Gemma (`gemma-4-26B`) and Mistral (`mistral-small-3.2`); out-of-family split rows for either family's numerators | `PREREG-S0.md:286` |
| The anchor's stand-in | After the anchor's closure, the next-oldest OpenAI id still callable; the change printed as a bridge row, both ids read on the same set in the reading of the change | `PREREG-S0.md:121`; the fixed ratio's frozen 88 at `:284` |
| The 96 GB card's blank | Every dollar cell `blank_reason: "P-6"` (no price, no dollar ratio, no dollars per anything) | `METRICS.md:118, :261`; `PREREG-S0.md:348, :353`; the rule's text `PLAN-v2.md:363` |

## The four later windows of reading 0.1

| Figure | Value | file:line |
|---|---|---|
| Reading 0.1's window end | 2026-10-02T08:19:34Z, the last arm of the 96 GB card's second window, stopped by its guard | the 96 GB card's second-window fold report, [`rtx-pro-6000-reading-0.1/CB5B-FOLD.md`](/the-cost-of-intelligence/data/rtx-pro-6000-reading-0.1/CB5B-FOLD.md), line 11 |
| The frontier on the docent gate (cells 2 to 6) | `gpt-4-turbo-2024-04-09` 58/63, `gpt-4o-2024-05-13` 59, `o3-mini-2025-01-31` 63, `o4-mini-2025-04-16` 62, `gpt-4o-mini-2024-07-18` 62; registered 36-page subset 34, 34, 36, 36, 35; 0 refusals; $7.6071; first rows 15:56:44Z–16:00:24Z, last batch 18:04:12Z, 2026-10-01 | the cells-2-to-6 report, [`openai-reading-0/receipts/cb1-part2/CELLS-2-6.md`](/the-cost-of-intelligence/data/openai-reading-0/receipts/cb1-part2/CELLS-2-6.md), lines 25–29 (the five scores, the 36-page subset, each cell's first row, refusals none), 36 (no refusal of any kind), 40 (the dollars) and 46 (the last batch's 18:04:12Z; also at 91 and 204) |
| The 96 GB card's first window | 2026-09-28, timer-opened 08:00:07Z, guard-stopped 08:39:22Z; 300 W; the card's main live service stopped, the others and a second card serving beside it under a guard; `llama3.3:70b` 93/108, `qwen2.5:72b` 93, `gpt-oss:120b` low 84, `llama3:70b` 71 (94 lenient, 27 late letters), `llama2:70b-chat` 1 (104 late, 42 lenient), `yi:34b` 19 (28 lenient, 40 off-form), `gemma2:27b` 72; both local anchors byte-identical to reading 0's beside-the-services rows on 312 items; 4 arms not run (a monitor's reader crash on a text-encoding fault; the guard at 39 minutes) | the first-window fold report, [`rtx-pro-6000-reading-0.1/CB5-FOLD.md`](/the-cost-of-intelligence/data/rtx-pro-6000-reading-0.1/CB5-FOLD.md), lines 3 (timer-opened 08:00:07Z; 300 W), 9 (guard-stopped 08:39:22Z; 39 minutes), 11 and 184 (the monitor's reader crash on a text-encoding fault), 20 and 51 (the 300 W cap held; the second board), 21 (the 70B scores and their lenient counts), 22 (both anchors byte-identical on 312 items), 46–49 (the four arms not run), 59–77 (the per-task rows; `gemma2:27b` at 77) and 103–106 (the anchors table); posture: `PREREG-S0.md` lines 522 (A-17 (1)) and 526 (A-24) |
| The 96 GB card's second window | 2026-10-02, 08:00:07Z–08:19:34Z; `phi4:14b` 82/108, `qwen3-30b-a3b-2507` 71, `qwen3.6:27b` 95, `gemma2:27b` 72 (byte-identical to its first window on 156 replies); `deepseek-r1:32b` reasoning on stopped by the latency guard after 8 items, recorded NOT RUN by the operators' ruling of 2026-10-02; the reference card's reading-0 row is the row of record | the second-window fold report, [`rtx-pro-6000-reading-0.1/CB5B-FOLD.md`](/the-cost-of-intelligence/data/rtx-pro-6000-reading-0.1/CB5B-FOLD.md), lines 11 (the window, 08:00:07Z to the stop at 08:19:34Z), 15–19 (the four scores; the fifth arm stopped by rule I-1), 22 (the 156 byte-identical replies), 24 and 26 (why I-1 tripped; the 8 items read) and 55–63 (the per-task rows); the NOT RUN ruling as quoted at [`laptop-reading-0.1/CB7-FOLD.md`](/the-cost-of-intelligence/data/laptop-reading-0.1/CB7-FOLD.md), line 224 |
| The laptop's reading-0.1 block | 2026-10-01, 19:09:33Z–20:06:45Z; 150 W on all 2,846 samples; `gemma2:27b` 74/108, `llama3.1:8b` 43, `mistral:7b-v0.3` 19, `mistral-nemo:12b` 46; `deepseek-r1:32b` reasoning on 1/108 strict, 86 lenient, SUPERSEDED (a registered memory check grew during the arm: replies stand, row not of record, energy enters nothing; the ruling of 2026-10-02); both local anchors 84 and 10, byte-identical to this card's reading 0 on 312 items | the laptop block's fold report, [`laptop-reading-0.1/CB7-FOLD.md`](/the-cost-of-intelligence/data/laptop-reading-0.1/CB7-FOLD.md), lines 5 and 11 (the window, 19:09:33Z to 20:06:45Z; 150 W on all 2,846 samples; the memory check), 24 (the memory check's count), 27 (both anchors byte-identical on 312 items), 28 (`deepseek-r1:32b`, 1 strict and 86 lenient), 59–71 (the per-task rows), 92 (the R-20 receipt row) and 136 (superseded: the replies stand, the row is not of record) |

## The human line ("Who ran this, and thanks")

| Figure | Value | file:line |
|---|---|---|
| The ids newest as of April 2026 | `claude-sonnet-4-6` released 2026-02-17, 98/108, $0.00344925 per correct (API-equivalent at list price, scaffold removed; not money paid); `claude-opus-4-7` released 2026-04-16, 97/108, $0.00210982 | `series.public.csv` rows: (`claude-sonnet-4-6`, the command-line tool, six-way, default effort), (`claude-opus-4-7`, the command-line tool, six-way, default effort), columns `weights_release_date`, `strict_correct`, `usd_per_correct_today` |
| The newest id at the reading | `claude-opus-5-5` released 2026-09-22, 104/108, $0.0031292 | `series.public.csv` row: (`claude-opus-5-5`, the command-line tool, six-way, default effort), the same columns |
| The drafting model's row | `claude-fable-5-1` 95/108 | `series.public.csv` row: (`claude-fable-5-1`, the command-line tool, six-way, default effort) |