Skip to content

Repository files navigation

shipshit.dev benchmark

A developer benchmark for frontier model releases. Every time a new model ships, the same task suite runs through the CLI coding agents people actually use (Claude Code, Codex CLI, Grok Build, Cursor), on ordinary subscriptions. Runs are scored on deterministic gates, hidden tests and blind rubric judging, with tokens, turns, wall time and API-equivalent cost next to every score. Results are JSON in this repo and a static site on Vercel.

Live site: https://benchmark.shipshit.dev (deployed by Vercel from master). Design and methodology: DESIGN.md.

Layout

apps/web           static site (leaderboard, runs, compare, methodology)
packages/schema    zod types: task.yaml, run results, judgments, releases, prices
packages/harness   `bench` CLI: run, judge, aggregate
packages/adapters  claude / codex / grok / cursor headless adapters
tasks/             the suite, one directory per task
data/              prices.json, runs/<run-id>/, releases/<release>.json

Commands

bun install
bun run bench run --release v2026.09 --agents claude:fable@high,codex:gpt-6-astra@high --tasks all --repeat 3
bun run bench judge --release v2026.09
bun run bench aggregate --release v2026.09
bun run --filter @benchmark/web build

Runs are idempotent on (release, task, agent, attempt); rerunning resumes.

bench validate checks every tasks/**/task.yaml against the schema and confirms the referenced files exist. bench list prints the suite, the supported CLIs and the configured judge panel.

Running a release

A bench run executes coding agents on subscriptions and needs every CLI installed and logged in on the machine that runs it (claude, codex, grok, cursor-agent), plus Playwright's Chromium (bunx playwright install chromium). Results land in data/runs/ and data/releases/; commit them and the site rebuilds on push.

Development

bun run lint                                  # biome
bun run typecheck && bun run test             # per package, through turbo
bun run --filter @benchmark/web sample:generate && bun run --filter @benchmark/web build:sample

The sample build renders a clearly synthetic release from apps/web/fixtures/sample-data/; nothing synthetic is ever written under data/.

License

MIT

About

Developer benchmark for frontier model releases: UX/UI, frontend, backend architecture, planning audit. Runs CLI coding agents on subscriptions, scores them, publishes results.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages