The rules are quite literal.

The joke is subjective. The chips aren't.

For each benchmark, subtract a model's score from the highest sourced score in the current snapshot. All scores here read higher as better.

What counts

We use public leaderboards and vendor reports. Each number links to its source and carries the date we observed it. Vendor results are labeled. We keep benchmark versions, tool use and context sizes separate. Missing results get no verdict.

This is a small snapshot, not an exhaustive leaderboard. Being near the top of this set does not mean being near the top of every model ever tested. Scores from different evaluation setups are imperfect comparisons. Read the conditions before treating a small gap as meaningful.

What the chips mean

They describe performance on one test. They don't promise reliability, safety, value for money or good judgment. We never average these scores into an overall rank.

What we will never do

No paid verdicts. No affiliate funnel. No signup, waitlist or subscriber email. No invented numbers or automatic refresh. An owner reviews sources and imports each update by hand.