Chat leaders
currentGemini 3 Flash
AR · May 13, 2026 · aggregate score 80.1 across 2 chat evidence rows.
Open modelHome · editorial front
The story behind
the AI rankings.
A reading mode for people who want the story behind AI rankings: what changed, what matters, and where the evidence is thin.
Data version
Read this before trusting a headline.
Data version May 13, 2026Model list checked9 providers · 800 tracked modelsPage refreshed May 18, 2026
If this date looks stale, you may be seeing an older build or cached deploy.
Issue 04Unbiased AI BenchEditorial front
Editorial front
Public AI rankings need
a more literate interface.
The point is not to crown one model. The point is to read the record: what was measured, by whom, under which judge, against which comparable group, and how stale the evidence already is.
Operating rules
1
Every score links back to a source page, raw value, benchmark version, and date.Bias becomes easier to inspect when the system refuses to flatten unlike things together.
2
Percentiles only compare scores from the same test setup.Bias becomes easier to inspect when the system refuses to flatten unlike things together.
3
Coverage gaps stay visible instead of being quietly filled in.Bias becomes easier to inspect when the system refuses to flatten unlike things together.
4
Parser anomalies and mapping fixes stay in the changelog.Bias becomes easier to inspect when the system refuses to flatten unlike things together.
Coding leaders
currentDeepSeek Reasoner
#1DeepSeek Reasoner81.7%
#2Claude Opus 4.773.5%
#3DeepSeek Chat63.4%
LB · Feb 6, 2025 · aggregate score 81.7 across 3 coding evidence rows.
Open compareFreshest source
opsTerminal-Bench
Loaded 28 Terminal-Bench 2.0 benchmark records from verified rows.
Open sourceA leaderboard without its measurement context is just a stronger-looking opinion. This product keeps the context on the page.
Method
Why percentiles only compare like with like
We normalize only when the underlying unit, judge, and benchmark version actually line up.
Read methodology →Compare
Head-to-head beats universal ranking when the surface is uneven
Comparisons stay grounded in shared coverage, raw values, and visible gaps instead of a universal scalar.
Open compare →Operations
Changelog entries matter because data plumbing changes outcomes
Parser fixes, mapping corrections, and source updates change what appears true. They need their own paper trail.
Open changelog →