The cheapest model that passes, by kind of task (run of 2026-11-04)
The pass rule, stated before the results: a model passes a kind of task only if it answered every question of that kind right, in all three repeats. Among the models that pass, the pick is the one with the lowest cost per 100 answers in that kind.
| Kind of task | Cheapest model that passes | Its cost per 100 answers | Did not pass (right/asked) |
|---|---|---|---|
| dates-units | GPT-6 Luna | $0.0026 | GPT-4o mini (9/12), GPT-4.1 mini (9/12), GPT-4.1 (9/12) |
| extraction | GPT-4o mini | $0.0024 | none |
| instructions | GPT-6 Luna | $0.0024 | GPT-4o mini (12/15), GPT-4.1 mini (9/15) |
| knowledge | GPT-4o mini | $0.0004 | none |
| reasoning | GPT-4o mini | $0.0006 | none |
Cost per 100 answers, by kind and model
| Model | dates-units | extraction | instructions | knowledge | reasoning | Cost per 100 answers, all kinds |
|---|---|---|---|---|---|---|
| GPT-6 Luna | $0.0026 | $0.0026 | $0.0024 | $0.0008 | $0.0018 | $0.0020 |
| GPT-6.1 Terra | $0.0137 | $0.0158 | $0.0234 | $0.0044 | $0.0133 | $0.0143 |
| GPT-6.1 Sol | $0.0255 | $0.0313 | $0.0334 | $0.0087 | $0.0229 | $0.0245 |
| GPT-4.1 | $0.0067 | $0.0304 | $0.0090 | $0.0055 | $0.0084 | $0.0113 |
| GPT-4o mini | $0.0005 | $0.0024 | $0.0007 | $0.0004 | $0.0006 | $0.0009 |
| GPT-4.1 mini | $0.0013 | $0.0062 | $0.0018 | $0.0011 | $0.0017 | $0.0023 |
What follows from the table
- In extraction, knowledge and reasoning every model passed, so the pick is simply the cheapest: GPT-4o mini for extraction, GPT-4o mini for knowledge and GPT-4o mini for reasoning.
- Dates and units is where the pick changes: only GPT-6.1 Terra, GPT-6.1 Sol, GPT-6 Luna passed, and the cheapest of them is GPT-6 Luna at $0.0026 per 100 answers.
- One model for everything: the cheapest model that passed all five kinds is GPT-6 Luna at $0.0020 per 100 answers.
- The money is small: the gaps between the picks are fractions of a cent per 100 answers, so at modest volumes the pass rule (and the cost of an occasional wrong answer) matters more than the price.
How the run was made (2026-11-04)
25 small tasks in five kinds (dates and units 4, extraction 4, instructions 5, knowledge 4, reasoning 8), each asked 3 times, so 75 answers per model. Same prompt to every model, temperature 0, no system prompt. A program scores every answer (a number within its tolerance, an exact word, JSON equal to the expected value, or a format rule such as a line count, word count or forbidden letters); no model judges another. A wrong format counts as wrong. No request failed, so every model has 75 scored answers. Repeats at one temperature are not independent trials, so the three repeats do not triple the evidence. Seconds are wall time from one PC and include the network. Prices are the providers' list prices as the run read them on 2026-11-04 (they may have changed since); the cost of a model is its input and output tokens times those prices.
What this run cannot tell
It cannot say how a model does on tasks that are long, open-ended or in another language, only on these 25 short ones. 25 tasks is a small sample: a model that scored 100% here can still miss on a task nobody wrote. It cannot say how prices or models change after 2026-11-04. The tasks are published as a list of kinds, and the run is repeated on a date so a change can be seen.