---
license: "mit"
language:
- en
pretty_name: "Cost Per Task"
tags:
- benchmark
- llm-evaluation
- ai-benchmark
- agent-benchmark
- agent-evaluation
- model-evaluation
- cost-per-task
- llm-cost
- evals
- grading
- openrouter
- openai
- mcp
- typescript
task_categories:
- text-generation
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: test
    path: data/tasks.jsonl
---

# Cost Per Task

Dollars per verified result, not tokens per second.

## What it measures

Cost, retries, turns, and tokens required for deterministic back-office task success.

Category: `agents`. Status: `alpha`.

## How to run

```console
$ npx benchmark-cost-per-task list
$ npx benchmark-cost-per-task run --model <model-id> --provider openrouter
$ npx benchmark-cost-per-task score results.json
```

`--provider mock` runs the whole benchmark offline with no keys. Real models run
through OpenRouter (`OPENROUTER_API_KEY`) or any OpenAI-compatible base URL
(`OPENAI_BASE_URL`, `OPENAI_API_KEY`).

## Data fields

`data/tasks.jsonl` holds 30 tasks, one JSON object per line
(kinds: 30 episode).

| field | type |
|---|---|
| `id` | string |
| `kind` | string |
| `prompt` | string |
| `input` | object |
| `expected` | object |
| `metadata` | object |

## Provenance

# Cost Per Task provenance

This file is generated from the benchmark definition.

## Seeded back-office episodes

- Kind: `synthetic`
- URL: https://github.com/pooriaarab/benchmark-cost-per-task/blob/main/scripts/gen-tasks.mjs
- Licence: `MIT`
- Retrieved: 2026-10-15T00:00:00.000Z
- How to get it: Run `node scripts/gen-tasks.mjs --seed 20261015 --split public` from the repository root.
- Contamination notes: The public tasks and generator are visible. Use `--split hidden` with an unpublished seed for final evaluations.
- Refresh: Generate and review a new public seed every six months.

## OpenRouter model price snapshot

- Kind: `live-feed`
- URL: https://openrouter.ai/api/v1/models
- Licence: `NOASSERTION`
- Retrieved: 2026-10-15T08:04:45.790Z
- How to get it: Run `node scripts/fetch-prices.mjs`; no API key is required.
- Contamination notes: Factual pricing metadata, not evaluation content; OpenRouter states no separate data licence for the endpoint.
- Refresh: Refresh before releases and at least monthly; retain git history for audits.


## Grader

Deterministic replay followed by exact virtual file-system and CRM state verification.

## Limitations

This dataset contains only the public task split (30 tasks).
Hidden evaluation splits use a private seed and are deliberately not part of
this dataset. Source kinds: live-feed, synthetic. These are evaluation prompts
for the `benchmark-cost-per-task` benchmark, not training data; scoring requires the package's
deterministic grader. See PROVENANCE.md for contamination notes.

## Citation

```bibtex
@misc{benchmark_cost_per_task,
  author = {Pooria Arab},
  title = {Cost Per Task task data},
  year = {2026},
  url = {https://github.com/pooriaarab/benchmark-cost-per-task}
}
```

## Links

- npm: https://www.npmjs.com/package/benchmark-cost-per-task
- GitHub: https://github.com/pooriaarab/benchmark-cost-per-task
- Leaderboard: https://aibenchmarkindex.com/benchmarks/benchmark-cost-per-task
- Dataset: https://huggingface.co/datasets/pooria/benchmark-cost-per-task