Free tool

Agent benchmark cost per solved task: $0.0468 or $234.2405

Pick a harness and model from the November 2026 boards. The page divides the listed run cost by the tasks that run resolved. Then it multiplies that rate by a task count you type.

The rows were read November 5, 2026. They are the same file as the agent harness benchmark, November 2026. A cell with no cost stays Not listed.

Loading the calculator.

What does one solved task cost on these boards?

$0.0468 on the cheapest positive SWE-bench Verified row, and $234.2405 on the dearest Terminal-Bench 4.0 row.

mini-SWE-agent with DeepSeek V3.2 Reasoner, result date 2025-12-01, resolved 60% of 500. The file lists the run cost as $14.04. Divided by 300 resolved issues, that is $0.0468 per solved task. The dearest positive row on the same mini-SWE-agent list is GPT 4o, $765.57 for 108.1 resolved issues, $7.0821 each, at 21.62% of 500.

Terminal-Bench is a different test, and the harness changes by row. Claude Code with Sonnet 5, dated 2026-06-30, lists $9,603.86 for 41 / 330. $9,603.86 divided by 41 successes is $234.2405. The cheapest positive listed rate on that board is Claude Code with GLM-5.3-Flash: $118.74 for 118 / 330, $1.0063 per success. A listed $0.00 stays $0.00 and is not treated as the cheapest positive rate.

Which rows stay Not listed?

Every team-checked SWE-bench row that is not mini-SWE-agent, and the one Terminal-Bench cost the board marked partial.

30 of 30 rows on the team-checked table have no costUsd in the file read November 5, 2026. Their cost per solved task is Not listed. Grok Build with Grok 4.7 prints $3,683.29 (partial: 324/330 trials) That partial figure is not divided and not scaled up to 330 trials. The harness benchmark copies the same cells and does not fill a blank.

What changes when you project N tasks of your own?

The listed rate times N, rounded to the cent. A missing rate stays Not listed.

At N = 100, DeepSeek V3.2 Reasoner is $4.68 and Sonnet 5 is $23,424.05. Change N and both move. The product assumes your task costs what one resolved issue, or one Terminal-Bench success, cost on that row. It does not know your repository, your retries, or a model the file does not list. A token bill from published per-token prices, which is a different meter, is the agent cost calculator.

FAQ

What is the cost per solved task on these boards?

On SWE-bench Verified with mini-SWE-agent, DeepSeek V3.2 Reasoner lists $14.04 for 300 resolved issues, $0.0468 per solved task. On Terminal-Bench 4.0, Claude Code with Sonnet 5 lists $9,603.86 for 41 successes, $234.2405 each. The file was read November 5, 2026.

How is tasks resolved calculated?

SWE-bench rows store a percent. Tasks resolved are that percent times the instance count in the file: 500 for Verified and 300 for Multilingual. Terminal-Bench rows store successes and trials, and the division uses successes. A percent that does not land on a whole number stays at one decimal. This page does not round it to a hidden integer before dividing.

Why do some rows say Not listed?

30 of 30 team-checked SWE-bench rows outside mini-SWE-agent have no run cost in the file. 1 Terminal-Bench row prints a partial cost, Grok Build · Grok 4.7: $3,683.29 (partial: 324/330 trials). Those rates stay Not listed. A listed $0.00 stays $0.00.

Does the projection estimate a missing cost?

No. N times the listed cost per solved task is a projection of that listed rate. If the row’s cost is missing or partial, the projection is Not listed. The number is not a quote for your repository, and it is not a fill-in for a board cell the file left blank.

Is what I select stored?

No. The calculation stays in the browser. The share link puts the run and N in the address bar. This site does not receive the form, and there is no email field.