Snazzyham's AI Benchmark

Models tested against my personal workflows, rather than an arbitrary benchmark.

Click a row for its task scores, a column header to sort.

Digest

Compress a 3,853-line business-rule spec and a 723-line client audit, in under 40 lines each.

What the tasks check
D1 checks a business-rule specification is compressed without losing conditions
D2 checks a client audit is compressed without losing critical issues
ModelScoreCostTimeTokensRun

Plan

Name every file a known change has to touch. The real fixes touched 25 and 19 files.

What the tasks check
P1 checks a plan names every file affected by the icon-system change
P2 checks a plan names every file affected by the materialised-order-cost change
ModelScoreCostTimeTokensRun

Build

Ship a CMS section end to end, and apply a design batch that then gets corrected mid-task.

What the tasks check
B1 checks a CMS section is shipped end to end and correctly registered
B2 checks a design batch is applied and corrected after a steering request
ModelScoreCostTimeTokensRun

Review

Three storefront-header and cart/menu diffs: the header change with two injected requirement violations, the same header change authentic, and a small correct mobile fix.

What the tasks check
R1 checks a storefront-header diff for the two injected requirement violations
R2 checks that a storefront-header diff is correctly recognised as clean
R3 checks that a correct mobile navigation fix is recognised as clean
ModelScoreCostTimeTokensRun