Datasets:

Mihai-LeanZero
/

leanzero-agentic-benchmarks

license: cc-by-4.0
pretty_name: LeanZero Agentic Benchmarks
language:
  - en
tags:
  - benchmark
  - leaderboard
  - agents
  - code-generation
  - llm-evaluation
  - atlassian-forge
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/leaderboard.jsonl

LeanZero Agentic Benchmarks

Aggregated results of LeanZero's agentic coding benchmarks, exactly as ranked on leanzero.net/agentic-benchmarks. One row per model per published benchmark run (a Forge 2.0 cloud baseline row covers its independent builds). Synced automatically from the site's data; the site is the source of truth and every row links back to its run page there, with the checks, screenshots, video and cost.

Leaderboard view: Mihai-LeanZero/leanzero-agentic-leaderboard.

Current contents: 115 rows (Gauntlet 7.2: 49, Forge 2.0: 35, Forge 1.0: 31).

The benchmarks

The descriptions below are the site's own wording.

Gauntlet (board, methodology). "AI models build a payments app. We test what works and record the cost." "Swarm and single-model agents build frozen app specifications and are graded through backend checks and browser evidence. Compare within the same scorer version." Grading: "Execution-based grading: the built application is started, probed over HTTP and driven in a headless browser, check by check". This dataset holds the current version, Gauntlet 7.2 (sb-7.2).

Forge (board). "AI models build a Jira app. We test the results and record the cost."

Forge 1.0 (forge-1.0, frozen): "The agent builds an Atlassian Forge app for Jira offline from a frozen contract; the scorer lints it, runs it in the Forge runtime against a seeded Jira site, drives every surface in a browser, and records per-check evidence and screenshots."
Forge 2.0 (forge-2.0, current): "one model ships v2 of an Atlassian Forge app for Jira offline, then it is run and graded." "Models upgrade the same Jira app in two independent builds, each tested and scored, with recorded inference spend shown for both."

Versions are not comparable with each other, and Gauntlet is not comparable with Forge: "a different task and a different scorer".

Method and ranking

Every score is from 0 to 1, produced by the benchmark's scorer, which runs the built app. A row is one published run, so run-to-run variance is not averaged out (except Forge 2.0 cloud baselines, below).
Gauntlet 7.2 and Forge 1.0 rank by score, highest first.
Forge 2.0: "Cloud baselines show the median and range of independent builds within the same verified task, engine and captured comparison cohort." They rank by that median; build_scores holds each build's own score and a single build is marked provisional. Other runs rank by their own score in the same list.
rank is the position on the leanzero.net board at sync time. The export runs the site's own ranking code on the site's own data, so the order matches the site. Rows with equal scores are equal; their order between them carries no meaning.
Cost: "From the saved OpenRouter generation billing summary". cost_status incomplete means "This is a lower bound; the billing record is incomplete." unavailable means no dollar amount was retained. "Scoring reuses the saved build; this amount describes the original inference spend." A row's cost_usd is the sum over its builds, or null when any build has no amount.
cost_source says where the amount comes from: openrouter-generation-records is a provider bill; declared-prices (with cost_status estimated) is an ESTIMATE the publisher made from their own declared USD per 1M token prices and the publisher-reported token counts, used when no provider bill exists. It is not a bill. mixed means the row's builds differ; a row with any estimated build is an estimate as a whole. null = no amount.

Fields

field	meaning
`benchmark`	`Gauntlet` or `Forge`
`version`, `version_name`	scorer version (`sb-7.2`, `forge-1.0`, `forge-2.0`) and its display name
`current_version`, `version_frozen`	whether this version is the family's current one, and whether it is closed
`rank`	position on the leanzero.net board
`model_id`	the model id the run used
`model_name`	the board's display name
`provider`	the API route the run used (`openrouter`, `google`, or `openai` for any OpenAI-compatible endpoint, which for a local run is the local engine)
`entrant_class`	`cloud baseline` (run by LeanZero), `community run`, `community fleet`, or `LeanZero fine-tune (local)`
`score`, `score_kind`	the ranked score and what it is (`single build`, or `median of N builds`)
`build_count`, `build_scores`, `score_min`, `score_max`	per-build scores
`tier_scores`	per-tier scores (median across builds where there are several)
`cost_usd`, `build_costs_usd`, `cost_status`, `cost_source`	inference spend in USD, and whether it is a provider bill or a declared-price estimate
`tokens_prompt`, `tokens_completion`, `tokens_cached`, `tokens_reasoning`, `tokens_total`, `token_calls`, `token_missing_calls`	the builds' token counts summed (every model call: the entrant's and the harness's helper calls); empty when any build lacks that figure. `tokens_total` is prompt + completion; reasoning is as the provider reports it and is not added
`token_source`, `tokens_exact`	where the counts come from (`provider-records` = the provider's own per-call records, e.g. OpenRouter; `session-counters` = the API usage goose summed; `engine-telemetry` = a local engine's own counts; `estimate`; `mixed`), and whether every build's counts are exact (false = a lower bound or an estimate)
`date`	the day the (first) build started
`board_url`, `result_url`, `build_urls`	links to leanzero.net

What is NOT here, on purpose: prompts, transcripts, generated code, file contents, screenshots, video, per-check evidence and anything that identifies who posted a run. Those stay on leanzero.net (the evidence is on each run page).

Licence and citation

CC-BY-4.0. Please credit "LeanZero, https://leanzero.net/agentic-benchmarks" and link the run pages you cite.