GPT-4o mini is the cheapest model for tasks

On the 2026-11-04 run, the rule was strict: a model only passes a kind of task if it answered every question in that kind correctly, across all three repeats[1]. For each kind, the pick is the lowest-cost model that met this bar[1]. Below are the threshold costs and picks for each task type, with the detail and where this answer would change.

The pass rule: strict correctness on every question

A model only passes a task kind if it answers every prompt in that kind correctly—one miss in any repeat and it is out[1]. We applied this rule to 25 short, English-language tasks (dates-units, extraction, instructions, knowledge, and reasoning), each repeated three times for consistency; only fully correct answers, by the test's format and tolerance rules, counted as right[1].

Task breakdown:

• Dates-units (4 tasks)
• Extraction (4 tasks)
• Instructions (5 tasks)
• Knowledge (4 tasks)
• Reasoning (8 tasks)

Every model saw the same prompts, scored by a program that checked for exact matches (with format or number tolerances as stated)[1]. No answer was judged by another model; each wrong format counted as wrong. No request failed; all models gave a full set of answers[1].

Table: Model picks and costs by task kind

The table below summarizes which models passed each kind and their cost per 100 answers, as measured on 2026-11-04[1]:

On extraction, knowledge, and reasoning, every model passed, so the pick is the cheapest: GPT-4o mini[1]. For dates-units and instructions, only GPT-6 Luna (and more expensive models) passed; no lower-cost model got every answer right[1].

Where the pick changes: dates-units and instructions

For dates-units and instructions, some cheaper models (GPT-4o mini, GPT-4.1 mini, GPT-4.1) missed at least three answers out of twelve or fifteen, so they do not pass under this rule[1]. The only passing models for these categories were GPT-6 Luna, GPT-6.1 Terra, and GPT-6.1 Sol, with Luna as the lowest-cost at $0.0026 for dates-units and $0.0024 for instructions per 100 answers[1].

If your use case requires date arithmetic, unit conversion, or precise instruction-following, you will need GPT-6 Luna or a more expensive model to guarantee a perfect record on tasks like those tested[1]. For extraction, knowledge, or short reasoning tasks, GPT-4o mini suffices at a much lower cost[1].

One model for everything: the all-pass pick

If your workflow needs a single model that will guarantee correct answers across all five kinds of short tasks, only GPT-6 Luna passed every one in this run, at $0.0020 per 100 answers[1]. No lower-cost model got every answer right in every category[1]. For most tasks, this is a premium over GPT-4o mini, but it is required if you cannot risk even a single error in any kind[1].

If you can route tasks by kind, you can mix: send extraction, knowledge, and reasoning to GPT-4o mini where it costs no more than GPT-6 Luna ($0.0024, $0.0004 and $0.0006 against $0.0026, $0.0008 and $0.0018 per 100 answers), but keep dates-units and instructions on Luna[1]. But for a one-model, error-free pipeline, Luna is the threshold.

What would change the picks: price, models, or tasks

These picks are from a single dated run (2026-11-04) using the models and prices available that day[1]. If prices drop or new models appear, the cheapest passing model could change—especially if a cheaper model is upgraded to pass the strict bar on dates-units or instructions[1]. The picks are also tied to the specific, short, English-language tasks tested; a different set of tasks or a different language could yield a new result[1].

If your volume is modest, the absolute cost differences are fractions of a cent per 100 answers, so the stricter pass rule and the risk of a wrong answer matter more than tiny cost gaps[1]. The table is a snapshot: rerun it with your own tasks, or check for newer price or model updates, if you need a current answer.

A worked example

Suppose you have a workflow that includes 1,000 instruction-following questions and 10,000 extraction tasks per day. You want to minimize cost while ensuring every answer is correct.

• For extraction, all models passed, so you can use GPT-4o mini at $0.0024 per 100 answers, totaling $0.24 for 10,000 answers.
• For instructions, only GPT-6 Luna passed, at $0.0024 per 100 answers, totaling $0.024 for 1,000 answers.

If you routed all tasks to GPT-6 Luna, your total would be $0.26 for 10,000 extraction answers ($0.0026 per 100) and $0.024 for 1,000 instructions answers, for $0.284. By routing extraction to GPT-4o mini and instructions to Luna, the total is $0.264 ($0.24 + $0.024), which saves $0.02 a day against $0.284, while meeting the strict correctness bar for both categories.

Questions readers ask

What if I only care about most answers being right, not all?

This run's rule is strict: every answer must be right to pass[1]. If you are comfortable with a small failure rate, you might use a cheaper model for more kinds, but the data here does not say how often or which questions would be missed. You would need a test with your own bar and error tolerance.

How much does the price difference matter at real scale?

The cost gap between models is a fraction of a cent per 100 answers[1]. At small scale, this is negligible; for millions of answers, it adds up. However, the cost of a wrong answer is often higher than the raw price difference, so the strict pass rule remains the decisive factor unless your volume is extremely high.

Will these results hold for longer or open-ended tasks?

No; this run only tested short, English-language tasks with precise, checkable answers[1]. Different models may perform differently on longer, more complex, or non-English tasks. The picks here do not generalize beyond the 25 tasks tested.

*Corrected 2026-11-05: The five example questions listed under the task breakdown were not the bench's tasks and are removed. The claim that GPT-4o mini costs under a tenth as much was wrong (it costs no more than GPT-6 Luna in those kinds, by the figures now given), and the worked example's all-Luna total is corrected from $0.244 to $0.284.*

Sources

1. Our own data, 4 Nov 2026: Model bench, 2026-11-04 (raw answers and scores)