The bench — what one correct answer costs, rented or made on a card you own, measured again every few months
The cost of intelligence, as a utility
exhibit eighty-five The bench
Published 2026-10-17 (UTC)
Reading 0 ran from 2026-09-23T15:40:04Z (its first row) to its registered close, 2026-09-25T12:00Z; reading 0.1, the first dated update, closed 2026-10-14T00:00Z
A small (human) team and a fleet of AI agents.
This page measures what one correct answer from an AI model costs, in dollars and in energy, rented from a company by the token or made on a graphics card you own, and how that cost moves over time: intelligence priced the way a utility is, by the unit delivered. It is a working document, the first page of a series measured again every few months. The figures below are reading 0, the series' first round: its measurements are frozen, and their scoring was corrected once, on 2026-09-27, to the form the pre-registration names ("What we got wrong" says how); neither is rewritten in later readings. Each later reading is added under its own dated heading further down the page, in the same shape: what was counted, the window it was counted in (UTC, both ends), and one sentence on how it compares with the reading before. A changes log at the foot lists every update. The first dated update, reading 0.1, is below.
ask about this page → assistant.strata2signal.com · in beta, still being tested
hardware on this page: RTX PRO 6000 Blackwell (96 GB) · RTX 3090 (24 GB) · RTX 5090 Laptop GPU (24 GB) → research.strata2signal.com/hardware/ · the roster is in beta, still being tested
the short version
This page measures what one correct answer from an AI model costs, in dollars and in energy, whether rented from a company by the token or made on a graphics card you own, and how that cost moves over time. The questions are 108 lines that two open AI models wrote in the voices of six historical dinner guests; the task is to name the speaker. The yardstick is OpenAI's June 2023 flagship, gpt-4-0613, at 88 of 108. The first open model that fits a 24 GB card and matches it, item by item with a margin of error, has weights published 2026-02-24: qwen3.5:27b, 97 of 108 on our reference card, an EVGA GeForce RTX 3090 XC3 Ultra, and 96 on a laptop's RTX 5090. OpenAI's price for a question at that level fell about 299-fold from June 2023 to September 2026, a factor of ten of it in one step. On the reference card a correct answer from that model takes 231 joules (a joule is one watt for one second), 71 of them in the seconds the server spent computing; most of the rest is a model-loading step this bench's server repeats on every question. At the yardstick's level the owned answer costs more than the cheapest rented one on the reference card even running around the clock (1.7 times, on a card price this reading did not re-read) and less only on the laptop running around the clock (0.72). Every owned figure is provisional for those reasons. Three questions this reading cannot answer, and the first dated update, are below.
5,956 words, about 27 minutes to read.
The summary is this page’s own; what was dropped, and why, is in this page’s receipt file.
What this page measures
A language model is sold two ways: rent answers by the token (a word or part of one) at a price the vendor sets, or buy a graphics card once, download an open model and make the answers yourself for the electricity. The cost of intelligence, on this page, is what one correct answer costs from each source, in dollars and in joules, on the same fixed questions, measured again every few months as the models improve and the cards stay the same. Three things are counted, each with a rule under "How to read the numbers": accuracy per dollar, accuracy per joule (a joule is one watt for one second, read from the card's own power counter) and the ratio over time (how close the best open model on an owned card comes to the frontier, with its uncertainty, reading by reading).
The words this page leans on. A correct answer is strict: the reply opened with the letter of the right option; naming the guest in a sentence counts wrong, because the products that use these models need the letter. The six-way is the main test: 108 lines from a public dinner-table transcript, six guests (Darwin, Einstein, Hypatia, Ibn Sina, Sagan and Socrates), which guest said this line. The anchor is the frontier of June 2023, OpenAI's gpt-4-0613; the local anchor is one open model at a pinned digest (gemma4:26b-a4b-it-q4_K_M), run on every card at every reading as its consistency check, with a weak second (mistral:v0.1, about 10 of 108). Registered means written down and timestamped before the result it governs was known; a cell is one registered measurement. A bar is cleared or not under a paired test on identical items; at the boundary means a bound sits within one answer of 108 of the line. A stand-in posture is any reasoning or tool setting other than the registered one; a ratio built on one prints "provisional". The cross-fit best is picked on half the questions and scored on the other half, both ways round, so "the best on a card" is not chosen after the fact.
Where the question comes from
In September 2026 Epoch AI, a research group that tracks the economics of AI, published "The plunging price of thought" (Luke Emberson and David Roodman, 2026-09-22, CC BY 4.0), whose main chart is titled "The price of artificial thought may have fallen faster than for any other transformative technology in history", a hedge this page keeps. For public benchmarks it finds the cheapest model that reaches a fixed score and what it charged per question, and reads that price falling about 12.67 times a year in real dollars over 2023 to 2026, with no standard errors, as it says. This workshop runs its own AI products on graphics cards it owns, on the thesis that a card bought once rides the model-efficiency curve for free as better open weights fit the same memory each year; the honest question is whether the rented kind rides it faster, since a vendor's price also falls with cheaper chips, batching and competition, and a result against owning must be publishable here. Theirs is a price on public benchmarks; ours is a cost on this workshop's own questions; the two meet only as rates.
The instrument
The cards. An EVGA GeForce RTX 3090 XC3 Ultra (24 GB), the reference card, at a 300 W power limit, 86 % of its rating; an NVIDIA GeForce RTX 5090 Laptop GPU under its maker's Dynamic Boost, which moves the power limit with the processor's draw (150 W on every sample of the rows tabled here); and an NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96 GB, read at its 420 W limit beside the live products it carries and, by a registered house rule, publishing energy only, no dollar figure.
The frontier, three ways. OpenAI's models through their API, 29 ids from June 2023 on; Anthropic's Claude models through the Claude Code command-line tool, 12 ids, the served model asserted on every call, dollars derived from the published API price list and labelled "API-equivalent at list price, measured through the consumer CLI" because the tool runs on a subscription; and 20 open-weight models served from Ollama's cloud shelf, most of them too large for a 24 GB card, "cloud-served open weights", never "frontier". The anchor is gpt-4-0613 (released 2023-06-13; $30 per million input tokens and $60 per million output, unchanged since launch); OpenAI's model list, read 2026-09-23T15:40:18Z, announced its retirement for 2026-10-23, after which the registered stand-in, gpt-4o-mini-2024-07-18 (73 of 108), takes over its re-read while the frozen 88 stays the fixed ratio's denominator.
The sets, four, frozen and hashed in the kit: the six-way, 108 lines from the public transcript of this workshop's dinner-party exhibit, the long table, written in the guests' voices by two open models, one Gemma (gemma-4-26B) and one Mistral (mistral-small-3.2), so every row from either family carries a home-team flag; the docent gate, 63 yes-or-no questions over pages of this site; the doorman's set, 48 visitor lines written to test the exhibit's doorman, never published; and the field exam, 40 rulebook excerpts. Reading 0.1 adds the judge set (119 rulings) and the long-form questions of RuleSage, this workshop's rules desk. The six-way went public on 2026-09-19; every open model and OpenAI id on reading 0's headline rows has a weights or creation date before it, claude-opus-5-5 (2026-09-22) is marked clean on its maker's stated training cutoff of June 2026, and the 15 rows the kit marks not clean enter no ratio. Every backend got the same prompt bytes, at temperature 0 with a fixed seed wherever the backend allows; reasoning models were read twice on the local cards, off inside 8 visible tokens and on inside a registered budget; the Claude tool's scaffold of about 430 tokens per call is subtracted from the derived dollars. Only the test items went to a cloud model, and the doorman's set left before its egress was ruled ("What we got wrong").
How to read the numbers
- Accuracy is strict, under one parser; the bar is a paired test. One parser, frozen under a hash with 48 self-test cases, scores every backend; an unparsed reply is never a choice, and a lenient reading (a late letter or a name) never enters a ratio. Every rate carries a 95 % interval resampled over the 16 dinners the lines came from. "Clears the anchor" means, on identical items, a difference of at least zero and a 10,000-resample paired lower bound at or above minus five percentage points (minus ten on each half of the set); the headline bar must be cleared on both halves (even and odd items in kit order).
- The ratio has two denominators. The fixed ratio divides a card's cross-fit best score by the frozen anchor's; the moving ratio by the cross-fit best closed-API model released in the 180 days before the reading's close. Both are chance-corrected (one in six subtracted from both sides) and print with their intervals; neither alone is the headline.
- Joules name their instrument. Board is the card's own counter, once a second, the only instrument compared across cards; each local row prints joules per correct answer over the whole window (what an owner's meter sees, waiting included) and over the compute phase alone (the seconds the server spent on the prompt and the answer, apportioned at the window's mean watts).
- Dollars at launch and today. A frontier row is its own token counts times the published price list, frozen per reading ("today's price" is the list read on 2026-09-23; launch prices come from archived pricing pages). A local row carries one owner, the buyer: the whole machine's replacement price (for the laptop, the price paid) over a fixed life, plus electricity at the national residential rate, at six hours a day or around the clock. Every reference-card dollar figure is provisional: its capital cost rests on a rig price whose card listing had been struck before the figure was registered ("What we got wrong" has the dates), and the rig had not been re-priced when this page was built.
Reading 0: the headline
Of the 37 open models read on the reference card (32 on the laptop card) from Ollama's library as 4-bit files in September 2026, the first that fits a 24 GB card and clears the registered test against the frontier of June 2023 (gpt-4-0613, 88 of 108) has weights dated 2026-02-24: qwen3.5:27b, read at a 4,096-token context, 8 tokens, reasoning off, 97 of 108 on the reference card and 96 on the laptop card (a model not read could have crossed earlier). It clears on both halves of the set on both cards (paired lower bounds +1.8 percentage points on the reference card, +1.0 on the laptop), as does the registered alternative, the half-year's star qwen3.6:27b (weights 2026-04-21; 96 and 95), chosen by a rule fixed before the re-score, so the crossing sits in 2026H1 without picking the winner from 37 candidates after the fact. Neither the anchor nor the first clearer is from a family that wrote the lines.
Why "clears" and not "scored at least 88". qwen2.5:32b (weights 2024-09-17) reads 89 of 108 on the reference card and 90 on the laptop, gemma3:27b (2025-03-01) 89 on both, and neither clears: they got different items right from the anchor, and across resamples of the 16 dinners the 2024 model's 95 % lower bound sits 6.5 percentage points below it.
When the test was fixed. The paired form was written into the bench's rules on 2026-09-23, the first day of measurement, because the plain wording of the bar ("matches gpt-4-0613's score on this task (within its interval)") read two ways once the anchor's own 88 was in; the ruling is stamped about 17:30Z, the rules file committed at 18:34:55Z. By then qwen2.5:32b and gemma3:27b had run on the reference card; every open model on a 24 GB card that clears on both halves ran after the commit, the first from 20:16Z. The rented rows came earlier: 56 of the 61 had started before the commit (all 29 OpenAI rows, the anchor's among them; all 20 cloud-served rows; 7 of the 12 Claude rows), 30 of the 34 rented both-halves clearers among them. On a raw count the crossing would date to September 2024; on the registered paired test, to February 2026. The scoring changed once more after the close, the resampling correction of 2026-09-27 ("What we got wrong").
The vendor's price, its own sentence. OpenAI's list price for one question at the anchor's level, per item at launch, fell from $0.008846 with gpt-4-0613 on 2023-06-13 to $0.0000296 with gpt-6-luna on 2026-09-14: 298.65-fold in 3.25 years, 5.76 times a year in nominal dollars; 328-fold, 5.93 times a year in real dollars (CPI-U, base month 2026-08), slower than Epoch's all-ages 12.67. The last step alone cut the price tenfold (a factor of 10.18 from gpt-5.6-luna); without it the line is 29.3-fold, 3.05 times a year nominal (32.1-fold, 3.15 real). The fall is almost entirely the input list price going from $30 to $0.10 per million tokens.
The ratio, with its spread and no verdict word. On the reference card the cross-fit best open model reads 1.114 against the frozen anchor (95 % interval 1.016 to 1.246) and 0.918 against the best closed model of the last 180 days (0.834 to 1.013, provisional), both chance-corrected; on the laptop card, 1.057 and 0.871, both provisional because its pick on one half, gemma4:31b, was read at a stand-in posture and is from a family that wrote the lines, so the table also prints them on the 57 lines that family did not write (1.109 and 1.072). The reference card's fixed ratio is written with its spread and never as a bare verdict: above the anchor on every seed tried, the lower bound about 1.012 to 1.019 across 21 seeds and both estimators, within one item of the line on about half of them.
The ladder
The same question asked of every backend at the anchor's bar: how many models were read, how many cleared, which cleared first by weights date, and which clears cheapest today.
| Backend | Read | Clear the anchor's bar (on both halves) | First clearer (weights date) | Its score, of 108 | Cheapest clearer, $ per correct answer |
|---|---|---|---|---|---|
| The reference card, a 3090 at 300 W | 37 | 5 (4) | qwen3.5:27b (2026-02-24) | 97 | qwen3.8:27b, $0.000190 ($190 per million correct answers; buyer, six hours a day; provisional) |
| The laptop card | 32 | 4 (3) | qwen3.5:27b (2026-02-24) | 96 | qwen3.8:27b, $0.0000877 ($88 per million; buyer, six hours a day; at the boundary, one half fails: provisional) |
| The 96 GB card, beside its services | 5 | 0 | — | Best: gemma4:26b 86 | No dollar is printed for this card |
| OpenAI API | 29 | 19 (15) | gpt-4-0613 (2023-06-13), the anchor itself: not a crossing date | 88 | gpt-6-luna, $0.0000333 ($33 per million) at launch and today |
| Claude, through the command-line tool | 12 | 10 (10) | claude-sonnet-4-5 (2025-09-29), the oldest id read: not a crossing date | 97 | claude-sonnet-5 at low effort, $0.000717 ($717 per million; API-equivalent at list price, measured through the consumer CLI) |
| Ollama's cloud shelf (open weights) | 20 | 11 (9) | mistral-large-3:675b (2025-12-02), of the 20 the shelf served | 91 | Plan-included; no metered figure |
Source: the kit's c3-ladder.csv and c3-tests.csv; "clear" counts the whole set, the bracket those that also clear on both halves.
The staircase
OpenAI's cheapest model at or above the anchor's bar, by release date, at launch price, in dollars per item; a step is drawn only where a newer model clears the bar more cheaply.
| Model | Released | $ per item at launch | Step factor | Note |
|---|---|---|---|---|
| gpt-4-0613 | 2023-06-13 | 0.008846 | — | The anchor; both halves |
| gpt-4-turbo-2024-04-09 | 2024-04-09 | 0.003037 | 2.91 | Provisional: fails one half of the set |
| gpt-4o-2024-05-13 | 2024-05-13 | 0.001436 | 2.12 | At the boundary; fails one half of the set; provisional |
| gpt-4o-2024-08-06 | 2024-08-06 | 0.000749 | 1.92 | At the boundary; fails one half of the set; provisional |
| gpt-4o-2024-11-20 | 2024-11-20 | 0.000711 | 1.05 | At the boundary; fails one half of the set; provisional |
| gpt-4.1-2025-04-14 | 2025-04-14 | 0.000585 | 1.22 | Both halves |
| gpt-5-2025-08-07 | 2025-08-07 | 0.000448 | 1.30 | Both halves |
| gpt-5.6-luna | 2026-06-23 (API created) | 0.000301 | 1.49 | Both halves |
| gpt-6-luna | 2026-09-14 (API created) | 0.0000296 | 10.18 | Both halves |
Source: the kit's c2-staircase-steps.csv and c2-staircase-rates.csv. Survivorship: four of the nine ids carried an announced retirement at the close, gpt-4-0613, gpt-4-turbo-2024-04-09 and gpt-4o-2024-05-13 for 2026-10-23 and gpt-5-2025-08-07 for 2026-12-11; the staircase is drawn over models that were callable at their time.
The two ratios
| Card | Denominator | k (models in the local pool / in the frontier pool) | Local, cross-fit | Frontier | Ratio, chance-corrected | 95 % interval (clusters) | Provisional | Label |
|---|---|---|---|---|---|---|---|---|
| A 3090 at 300 W | Fixed: the anchor | 23 / 1 | 0.889 | 0.815 | 1.114 | 1.016 to 1.246 | No | Written with its spread (above) |
| A 3090 at 300 W | Moving: closed models of the last 180 days | 23 / 14 | 0.889 | 0.954 | 0.918 | 0.834 to 1.013 | Yes | Interval contains 1; the panel includes CLI-measured Claude rows |
| A 3090 at 300 W | Cloud-served open weights (3 largest) | 23 / 3 | 0.889 | 0.861 | 1.040 | 0.925 to 1.122 | Yes | Interval contains 1 |
| A 3090 at 300 W | Moving, temperature-0 rows only | 23 / 6 | 0.889 | 0.954 | 0.918 | 0.834 to 1.000 | No | At the boundary |
| The laptop card | Fixed | 19 / 1 | 0.852 | 0.815 | 1.057 | 0.987 to 1.210 | Yes | At the boundary; home-team pick |
| The laptop card | Fixed, lines not written by the pick's family (57 items) | 19 / 1 | 0.935 | 0.860 | 1.109 | 0.918 to 1.226 | No | Home-team split |
| The laptop card | Moving | 19 / 14 | 0.852 | 0.954 | 0.871 | 0.815 to 1.000 | Yes | At the boundary; home-team pick |
| The laptop card | Moving, on the same 57 items | 19 / 14 | 0.935 | 0.884 | 1.072 | 0.881 to 1.176 | Yes | Home-team split |
| The laptop card | Cloud-served open weights (3 largest) | 19 / 3 | 0.852 | 0.861 | 0.987 | 0.905 to 1.102 | Yes | Interval contains 1 |
| The laptop card | Moving, temperature-0 rows only | 19 / 6 | 0.852 | 0.954 | 0.871 | 0.815 to 0.988 | Yes | Interval below 1; at the boundary |
| The 96 GB card | Fixed | 3 / 1 | 0.759 | 0.815 | 0.914 | 0.732 to 1.064 | No | Roster coverage, not the card (k = 3) |
| The 96 GB card | Moving | 3 / 14 | 0.759 | 0.954 | 0.753 | 0.608 to 0.870 | Yes | Roster coverage, not the card (k = 3) |
Source: the kit's c4-ratios.csv; "local, cross-fit" and "frontier" are strict accuracies before the one-in-six correction.
Energy, by instrument
| Model | Card | Score, of 108 | Mean board W | Board joules per correct answer, whole window | Board joules per correct answer, compute phase (apportioned) |
|---|---|---|---|---|---|
| gemma4:26b-a4b (the local anchor) | A 3090 at 300 W | 88 | 155.1 | 183.4 | 26.3 |
| qwen3.5:27b (the first clearer) | A 3090 at 300 W | 97 | 186.7 | 231.0 | 71.4 |
| qwen3.6:27b (the star) | A 3090 at 300 W | 96 | 186.9 | 233.6 | 75.1 |
| gemma4:26b-a4b (the local anchor) | The laptop card | 84 | 55.9 | 28.6 | 7.8 |
| qwen3.5:27b | The laptop card | 96 | 88.4 | 58.0 | 27.7 |
| qwen3.6:27b | The laptop card | 95 | 90.9 | 61.2 | 30.0 |
| gemma4:26b-a4b | The 96 GB card at 420 W, beside its services | 86 | 103.3 | 157.4 gross; 74.3 net of the services' idle draw | 12.7 |
Source: the kit's series rows and the 96 GB card's beside-the-services tables. Across the 67 cells both 24 GB cards read with a correct answer on each, the laptop card's whole-window board joules per correct answer are a median 0.27 times the reference card's and 0.42 times in the compute phase, the fairer pair because the two cards load a model at different speeds. Much of a whole-window figure is the server's per-question loading step: the local anchor's 108 questions took 104 seconds on the reference card, 82 of them, by the server's own timings, loading a model already in memory (on the laptop, 26 of 43 seconds), a step the card's report calls a property of this server, never to be carried to another.
The frontier's own curve
Through OpenAI's history the same 108 lines climb from 34 and 24 of 108 on the gpt-3.5-turbo snapshots of late 2023 and early 2024, through the anchor's 88, to 103 of 108 by gpt-5.6-sol, the flagships of 2026 bunched within a few answers of each other.
The Claude line, factually. Twelve ids read through the command-line tool at its default effort, one scored pass each (two ids twice, averaged); whether the account lets the vendor train on its traffic is not readable from the tool. From claude-sonnet-4-5 (released 2025-09-29) at 97 of 108 the line climbs, with dips, to claude-opus-5-5 (2026-09-22) at 104; claude-sonnet-4-6 (2026-02-17) reads 98 and claude-opus-4-7 (2026-04-16) 97; every id's row and date is in the kit. The twelfth, claude-fable-5 (2026-06-09), stands apart: its safety classifier refused 53 of the 108 dinner lines, every one labelled "cyber" content, each an ordinary line spoken by one of six historical guests. A refusal is a wrong answer under the registered rule, so that id scores 53 of 108; of the 55 lines it answered it got 53 right, a diagnostic that enters no rate. Its successor refused none.
The break-even that moves
What it asks. For a model that clears a bar on an owned card, at what share of the hours must the card be busy for the owned answer to cost less than the cheapest rented answer at the same bar? Reading 0 can only set the question up. The model behind every owned-dollar figure is qwen3.8:27b (weights 2026-08-05, clean): 95 of 108 on the reference card, clearing on both halves, and 93 on the laptop, failing one half; the cheapest clearer on both cards at six hours a day. The buyer's inputs are the machine's replacement price (for the laptop, the price paid) over a three-year straight-line life, electricity at the national residential rate for June 2026, and the seconds per answer the capital is spread over, the bench's task window divided by its 108 questions (1.03 s on the reference card, 0.55 s on the laptop); on the reference card about half of that is the server's loading step (a median 0.51 s), which no registered rule removes, and no figure paced on computing alone was computed, so no verdict word prints. The two prices are on the pages linked at the foot.
What reading 0 shows, provisionally. Against gpt-6-luna at today's price, the cheapest clearer on the reference card costs the buyer 5.70 times the rented answer at six hours a day and 1.68 times around the clock; on the laptop card, 2.63 and 0.72 times, the last under one and provisional because that clearer fails one half (qwen3.5:27b, which clears on both, reads 2.72 and 0.74). Every one of these figures is provisional: the reference card's capital rests on a struck listing ("What we got wrong"), and the laptop's fuel is its board's draw over an assumed 0.9 power-supply factor with the box's idle unmeasured, so it understates. The sensitivity rows (a five-year life, a cost of money, the commercial and industrial tariffs) are in the kit.
What is registered to be tested, and when. The pre-registration's one forward prediction: at the anchor's bar, the cloud's dollars per correct answer fall faster than the owned figure, by a factor between 1 and 1.6 a year. A reading moves the verdict toward ownership only when the owned-to-rented ratio has fallen by more than a registered band, against it only when it has risen by more, and otherwise prints "inconclusive"; the band's numbers were not registered at reading 0.1's close, the first reading that can print an outcome is reading 1, due 2026-12-23, and a yearly factor needs two readings a year apart. Until then this page prints the staircase's rates, 5.93 times a year in real dollars and 3.15 without the last step, and no direction against them.
What this page does not know yet
Three questions a reader asks, each with the cell that answers it.
- Does owning ride the curve as fast as renting? Not at reading 0: reading 1 (due 2026-12-23) can print the first outcome if the band's numbers are registered by then; a yearly factor needs readings twelve months apart.
- What does owning cost? Every local dollar and break-even figure is provisional until the rig is re-priced and the seconds per answer behind it are ruled.
- Does a cheaper card do it, and is it only this task? Whether a model that fits a 12 GB card clears the anchor's bar is a registered cell, as is a board-to-board control on a second 3090; neither had run by reading 0.1's close, 2026-10-14T00:00Z. What 96 GB buys, and how the anchor fares on the other sets, reading 0.1 began to answer, below.
What we got wrong on the way
A bench that only reports what worked is advertising. In order:
- A private set left this workshop's machines before its egress was ruled. The doorman's 48 lines went to the cloud shelf's models and, through the tool, to seven Claude ids (one by the tool's fallback, on 12 re-sent lines); the egress was held at about 17:38Z on 2026-09-23; none went to OpenAI.
- By default, the tool answered a refused question with a different model, and said so in its usage record; the served-model check caught it on the first day, every re-sent item left every rate, and the affected ids were re-read with that switch off: the 53 refusals above are the re-read.
- One parser now scores every backend. An earlier report of this instrument printed a chat model's score one answer high by counting an unreadable reply as the first option.
- The test was first computed on single lines, not on the dinners the pre-registration names. Under that form the laptop's first clearer was qwen2.5:32b; under the registered form, applied on 2026-09-27, it fails, and fourteen verdicts moved, eleven against the anchor.
- A model that does not stop thinking was first scored as if it had answered. deepseek-r1:32b's first pass printed 18 of 108 because empty replies were parsed as the first option; under the common parser that row reads 0 of 108.
- The laptop's power limit is a range, not a number. The bench gated each arm on one fixed wattage; under Dynamic Boost the firmware moves it.
- The reference card's dollars were registered on a price already struck. The whole-rig price behind them, read 2026-09-13, includes a card listing this workshop struck on 2026-09-18; the pre-registration of 2026-09-24 used it anyway.
- A registered prediction failed. The anchor was predicted to score under 50 % on the judge set; it scored 91 of 119 in reading 0.1, and the prediction stays on the page with the figure beside it.
- An announced window never ran in reading 0. The models too big for a 24 GB card were to be read on the 96 GB card before the close; they ran in reading 0.1 instead, above.
Reading 0.1: the first dated update
Reading 0.1 holds the cells not on reading 0's registered list, whenever they ran, so its window opens before reading 0 closed: 2026-09-23T15:48:50Z (the first Claude call on the doorman's set) to 2026-10-02T08:19:34Z (the last arm on the 96 GB card); it closed at 2026-10-14T00:00Z. Against reading 0 in one sentence: every figure of reading 0 stands as measured, the anchor reproduced itself on 106 of 108 items, both local anchors reproduced reading 0 byte for byte on the laptop card and on the 96 GB card, and reading 0.1 prints no ratio of its own (the last paragraph says why).
The first big-model window (2026-09-24, 06:35:55Z to 06:45:56Z). With every service on the 96 GB machine stopped, falcon:180b, split across two cards, put its letter late or as a word on every line: 0 of 108 strict, 19 lenient, a form result and not a capability reading.
The anchor, again (2026-09-27, 08:38:04Z to 09:47:17Z). gpt-4-0613 on the six-way again: 106 of 108 items chose the frozen row's letter, both passes 88, the paired difference 0.0 (−2.7 to +2.8 percentage points); the registered prediction of at least 105 agreeing items held. On the judge set, 91 of 119 (76.5 %) against a registered prediction of under 50 %, which failed; on the field exam, 40 of 40.
The anchor on the docent gate (2026-10-01, 15:10:04Z to 15:14:27Z). 24 of the 63 pages exceed its context window and score wrong, a capability failure shown as coverage: 39 of 63 strict, every page it was sent answered right, 36 of 36 on the registered subset that fits an 8,192-token window.
The Claude line at its lowest effort (2026-09-27, 08:21:55Z to 09:10:15Z). Three ids, three passes each: claude-opus-4-7 100.00 of 108 against its default row's 97, claude-opus-5 98.33 against 98, claude-fable-5-1 79.67 against 95, a drop of 15.3 answers (−23.7 to −7.4) that is form, not knowledge: its lenient mean is 101.0 and it refused none. The registered prediction, each low row within five answers of its default, held for two ids and failed for the third.
The anchor on the rules desk (2026-09-27, 08:49:49Z to 08:56:25Z). 60 cells of RuleSage's long-form questions, 54 collected, 6 refused for context length; the answers were sealed and not judged by the close, so no score prints.
The frontier on the docent gate (2026-10-01, 15:56:44Z to 18:04:12Z). Five more OpenAI ids on the same 63 pages: gpt-4-turbo-2024-04-09 58, gpt-4o-2024-05-13 59, o3-mini-2025-01-31 63, o4-mini-2025-04-16 62 and the bridge id gpt-4o-mini-2024-07-18 62; on the registered 36-page subset, 34, 34, 36, 36 and 35. None refused a page, so the anchor's 24 refusals were its window, not the pages.
What 96 GB buys, first window (2026-09-28, 08:00:07Z to 08:39:22Z, stopped by its guard). On the 96 GB card at 300 W, its other live services and a second card serving beside it under a guard, the open models too big for a 24 GB card, reasoning off (gpt-oss at its lowest effort): llama3.3:70b 93 of 108, qwen2.5:72b 93, gpt-oss:120b at low effort 84, llama3:70b 71 (94 lenient), llama2:70b-chat 1 (42 lenient); two half-year stars, yi:34b 19 and gemma2:27b 72. Both local anchors reproduced reading 0's beside-the-services rows byte for byte on all 312 items at a different power cap. Four arms did not run (a monitor's text-encoding crash, since fixed; the guard at 39 minutes).
The stars on the same card (2026-10-02, 08:00:07Z to 08:19:34Z). phi4:14b 82 of 108, qwen3-30b-a3b-2507 71, qwen3.6:27b 95, and gemma2:27b, the ruled re-read, 72, byte-identical to its first window on all 156 replies. A fifth arm, deepseek-r1:32b with reasoning on, was stopped by the latency guard after 8 items and is recorded not run, by ruling; the reference card's reading-0 row stays the row of record.
The laptop's four, the think-only star and both anchors (2026-10-01, 19:09:33Z to 20:06:45Z). Seven cells in one queue at 150 W on all 2,846 power samples: gemma2:27b 74 of 108, llama3.1:8b 43, mistral:7b-v0.3 19, mistral-nemo:12b 46; deepseek-r1:32b with reasoning on 1 of 108 strict and 86 lenient, in kind with the reference card's row, though a registered memory check grew during that arm, so by rule its row is not of record and its energy enters nothing; and both local anchors, 84 and 10, byte-identical to this card's reading 0 on all 312 items.
Reading 0.1's ratio. A reading's ratio prints only for a card whose two local anchors were re-read inside the reading and entered the series before the close. No card met that at 2026-10-14T00:00Z, so reading 0.1 prints no ratio on any set: the reference card's anchors were not re-read inside the reading, and the laptop card's and the 96 GB card's re-read rows (above) had not entered the series by the close; they can before reading 1 (due 2026-12-23), which could then print 0.1's ratio beside its own.
What to take with you
- The frontier of June 2023 fits a 24 GB card as of weights published 2026-02-24, on this task, among the models read: qwen3.5:27b, 97 and 96 of 108 against the anchor's 88, on a paired test, on both halves, on both cards.
- OpenAI's price for that level fell 298.65-fold in 3.25 years, a factor of ten of it in one step: 5.93 times a year in real dollars, 3.15 without that step; a price, not a cost.
- Against the frontier a 3090 reads 1.114 (1.016 to 1.246) on the 2023 anchor and 0.918 (0.834 to 1.013) on the last 180 days' best, written with its spread and no verdict word.
- On the bench's server, a large share of a correct answer's energy is the server's loading step, not computing: 231 board joules for the first clearer on the reference card, 71 of them computing.
- At the anchor's bar, owning is cheaper only on a laptop running around the clock, and the dollars are provisional: 5.7 times the rented answer at six hours a day on the reference card, 1.7 around the clock, 0.72 on the laptop card around the clock.
The case against owning, in one sentence: the vendor's price fell about 5.8 times a year at this bar, a factor of ten of it in one step an owner could not take, the owned dollar figures are worse than the rented ones on the reference card at every utilisation read, and no ratio here is shown rising (the laptop's moving ratio against the temperature-0 rows sits wholly below one: 0.871, 0.815 to 0.988, provisional). The case for, in one: the frontier of June 2023 fits a 24 GB card on this task within a factor of two of the vendor's cheapest answer at full utilisation, a rented price can also be withdrawn (OpenAI announced the retirement of three of the nine staircase models, the anchor among them, for 2026-10-23) while an owner's model file does not retire, and what a card buys that a meter does not, no per-call meter, no repricing, no words leaving the building, is not priced here.
How to check our work
- The kit. Every public row of every reading in series.public.csv, with its backend, set, posture, parser version, gates and sources, machine identifiers withheld; the four frozen sets by hash; the parser; the pre-registration, its amendments and the rules they fold, as public copies; the chart data as CSV; cut 2026-10-14 (UTC) and published beside this page.
- The receipts. Every figure on this page, the sentence it sits in and the kit file and line it was typed from: the receipts table.
- The sources. Luke Emberson and David Roodman (2026), "The plunging price of thought". Published online at epoch.ai. Retrieved from https://epoch.ai/publications/the-plunging-price-of-thought. Accessed 23 Sept 2026. Licence: CC BY 4.0. Epoch AI has not reviewed this comparison; its data are cited and linked, not redistributed.
- Run it yourself. A stand-alone runner (the six-way with the frozen parser against any local model server, one row, no energy) is planned for reading 1, due 2026-12-23.
The rest of the seminar
Related pages on this site: the long table itself, where the six-way's lines come from, A Dinner Party for the Dead; the parser lesson and the first cross-card read, Reading the answer instead of writing it and A laptop, asked the desktop's questions; the same model of card against its 3090 Ti sibling, RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench; the frontier models at this workshop's desks, Two New Frontier Models at the Rules Desk and Three New Voices at the Narrator's Chair; what the rig costs to run, read off a meter, What it costs to run a high-end AI rig at home; what it costs to buy, the source of this page's capital figure, The cost to purchase and run a very capable home AI rig; a second RTX 3090 of the same make and model under two serving engines, Ollama and vLLM on one RTX 3090; the licences the open models carry, The licence ledger; and two histories, A short history of Ollama and A short history of vLLM.
Who ran this, and thanks
A small (human) team ran the cards, paid the meters, read the receipts and ruled the open questions; a fleet of AI agents ran the measurements under a written workflow, one stream per card and one per vendor, each audited before anything merged. The humans on this team have worked with Claude models nearly every day since about April 2026. The newest ids then were claude-sonnet-4-6 (released 2026-02-17; 98 of 108; $0.0034 per correct answer at the API list price read 2026-09-23, not money paid) and, from the middle of that April, claude-opus-4-7 (97; $0.0021); the newest at this reading, claude-opus-5-5 (2026-09-22), reads 104 of 108 at $0.0031. The prose was drafted by a Claude model, Fable 5.1, also a measured row on the page (95 of 108). Epoch AI published the chart that started this page under a licence that lets us cite it; the open-weights makers published the models; Ollama's library hosted them. Thanks to all three, and to the six guests, who did not consent to any of this.
Changes
- 2026-10-17 (UTC) · reading 0 published: measurements closed 2026-09-25T12:00Z, scoring corrected to the registered form on 2026-09-27; the receipts table's line numbers are each file's in the bytes the kit's index.json pins by sha256.
- 2026-10-17 (UTC) · reading 0.1 added as a dated update, closed 2026-10-14T00:00Z; its four later windows landed in the kit with it.
This page, its chart data and its data kit are published under CC BY 4.0. Epoch AI's figure is cited under its own CC BY 4.0; its data are linked, not redrawn here.