Published by Electricity Studio
About Electricity Bench
Electricity Bench is a coding-agent evaluation project published and operated by Electricity Studio. The organization is responsible for suite design, grading, editorial decisions, publication, and corrections. We use organization-level attribution rather than personal bylines.
What we publish
We run coding-agent products through private task suites using the subscription plans and tools a customer would use. We publish scorecards covering capability, plan usage, task time, and reliability, along with comparisons, historical trends, methodology, and news.
Start with the leaderboard or read the methodology.
Reuse and license
Published scorecards, leaderboard tables, weekly run snapshots, methodology text, and news may be quoted and reproduced with attribution to Electricity Bench (an Electricity Studio project) and a link to the source page.
Private task prompts, fixtures, checker logic, and withheld attempt outputs are not published and are not licensed for reuse.
The Electricity Bench name and hex-badge mark remain Electricity Studio brand assets. Questions about reuse can be directed to Electricity Studio.
How grading works
The benchmark subject is the complete agent: its harness, selected model, settings, and plan tier. Agents run headless through their own tools. Suite versions pin their scoring method, task structure, repeats, and grade bands so results remain comparable within that version.
Task prompts and checker logic stay private to reduce memorization and benchmark contamination. The scoring method, grade bands, redacted outcomes, run provenance, and operational measurements are public.
Claims and commercial context
Provider claims do not contribute to Electricity Bench grades. When a relevant third-party number appears, it is kept in a separate context block with its source, retrieval date, and whether it is self-reported or independently produced.
Subscription prices and plan facts are sourced and dated. Plan quota percentages are never treated as the same unit across different providers, and a subscription run is never assigned a fabricated API price.
Corrections
Results are published as measured. When we find a run faulty - a harness failure, a leaked fixture, a wrong plan pin - we correct or withdraw it and note the change on the scorecard and the weekly run page it affects. Third-party claims never count toward a grade, so a correction never comes from a provider's own numbers.
Questions about a result or correction can be directed to Electricity Studio.
Privacy and analytics
Published data passes a redaction gate before it reaches this site. Private prompts, fixtures, checker logic, and unsafe attempt outputs are not exposed. Withheld material stays withheld even when a score is published.
Site analytics are first-party and cookieless. They record aggregate pageview context without storing IP addresses, and no third-party advertising tracker is embedded on the site.