aimpg scoreboard

Do token-savers, prompts, models and agents really save money on real code? Each result reruns a developer's own past commits both ways, and their own tests judge them. Made with aimpg verify; 8 records so far.

Report No. 1: what AI coding costs per solved task (984 pre-registered reruns of real commits) ยท aimpg in 5 minutes

Reproduced

Counted only when someone else reran a public record and got the same answer.

Nothing here yet: no claim has enough repos and people behind it.

Self-reported (not audited)

Nothing here yet: no claim has enough repos and people behind it.

Collecting data

claude-code@haiku vs claude-code

A number appears once at least 5 repos and 3 people back it, with no one supplying more than 50% of the runs. Each account counts for at most 3 repos per claim. Results use the median repo, never an average.

Disputed

Public records (rerunnable)

TierClaimAnswerByWeek
self-reportedclaude-code@haiku vs claude-codeSUPPORTEDkumarganduri2026-W42record
self-reportedclaude-code@haiku vs claude-codeNOT SUPPORTEDkumarganduri2026-W42record
self-reportedclaude-code@haiku vs claude-codeSUPPORTEDkumarganduri2026-W42record
self-reportedclaude-code@haiku vs claude-codeSUPPORTEDkumarganduri2026-W42record
self-reportedclaude-code@haiku vs claude-codeSUPPORTEDkumarganduri2026-W42record
self-reportedclaude-code@haiku vs claude-codeSUPPORTEDkumarganduri2026-W42record
self-reportedclaude-code@haiku vs claude-codeSUPPORTEDkumarganduri2026-W42record
self-reportedclaude-code@haiku vs claude-codeSUPPORTEDkumarganduri2026-W42record

This site counts page visits with GoatCounter: no cookies and no personal data. The aimpg tool itself sends nothing.