Benchmark data

Name: hibench coding-agent benchmark exports
Creator: hibench
License: https://opensource.org/licenses/MIT

Download the canonical hibench export tables used by the dashboard. Files are generated fromhibench export and copied into/data/ on each site build.

Benchmark data last updated July 7, 2026 UTC

Dataset overview

Schema version: hibench.benchmark.v1
Coverage: 16 agents · 1132 captured versions
License: MIT
Update cadence: Refreshed when benchmark captures are exported and the site is rebuilt

See the methodology for capture rules, tokenizer notes, and how primary requests are selected. Source repository:hibenchmark/hibench.

Downloads

runs.csv
One row per canonical primary capture: agent, version, tokenizer totals, tool/skill/MCP/sub-agent counts, and capture metadata.
tools.csv
Declared tools per primary request with definition size, type, and MCP/sub-agent flags.
skills.csv
Bundled skills per run with token cost and preview text.
subagents.csv
Declared sub-agent types per run with token cost and source classification.
mcp.csv
MCP server declarations per run when present.
text_fields.csv
Tokenized text-field breakdown (instructions, injected context, user prompt, tool context, and related categories).
export.json
Export manifest with schema version, row counts, and file inventory.
github_stars.json
Optional GitHub star metadata for agents with public repository links.