Benchmark Radar

Wu, Koutian

Today's radar

Daily briefing

Questions for today

Most reported benchmarks in model cards

Scores over time

Benchmark reported scores over time

How to read this chart

Every value that could be read verbatim from a cited document, placed at the date that document was published rather than at any evaluation date. The line directly links actual observations that set a new reported record among the values shown; it neither holds a score between reports nor extends past the final record. Test versions and run conditions can differ, so this is a reported-record path, not a like-for-like trend. When no newer number could be read, the gap is marked rather than drawn through. Whether a benchmark has saturated stays a reading you make, not a score this panel prints.

All tracked benchmarks

What would it take to chart best score against lowest cost?

Two scores can only be compared when each one records which version of the test it used, which slice of it, whether a high number is good or bad, which model, what software ran it, how much thinking time it was given, what it cost, how long it took, when it was published, and where the number came from. The test and the way it was run have to match; the model, cost, time and date are free to differ, because those are what the chart compares. This site records that a model card mentioned a benchmark. It does not yet record those measurements.

Once they exist, this chart could plot cost or speed against score, and draw a line through only the results that nothing else beats on both at once. A date slider would then show how that line moved over time.

Benchmarks by model card adoption

Each model card counts once per benchmark. A card reporting AIME in four configurations counts the same as a card reporting it once, so a long appendix cannot outweigh a different vendor. Organizations breaks the tie: the same count from six vendors is a shared standard, from one vendor a house style.

Audit the counts Model cards in the registry

The curated source list this ranking is computed from. Expand any card to see every benchmark it reports, grouped the way the source document groups them, so our data can be checked line by line against the original.

Big picture

What we found

See what shows up most often and where it came from. Open the connections view when you want to look at a specific item.

Want more detail? See how everything connects

There are a lot of dots here. Pick one to see what it connects to, or open the matching results.

Recent activity

Signals over time

Counts describe discovery volume, not scientific quality.

New by domain

Sep 1, 2026

benchmark

185

vs previous scan: +92
recent daily average: 104.67
vs its average: +77%
cumulative: 3,572
also updated (not counted above): 214

evaluation

119

vs previous scan: +85
recent daily average: 65.67
vs its average: +81%
cumulative: 2,774
also updated (not counted above): 121

dataset

81

vs previous scan: +36
recent daily average: 62.33
vs its average: +30%
cumulative: 2,313
also updated (not counted above): 108

agentic

29

vs previous scan: +9
recent daily average: 17.33
vs its average: +67%
cumulative: 785
also updated (not counted above): 40

data quality

3

vs previous scan: +3
recent daily average: 0.00
cumulative: 61
also updated (not counted above): 2

Daily evidence and attention volume

Category tags overlap. Each bar is an independent count, not a part of a stacked total.

Releases only Excludes records re-announced as an update to something already surfaced.

Dev checker

Source mix counts ranked evidence after scoring. Fetch health counts raw records returned before scoring, so a source can be ok and still empty.

Date	Coverage (UTC)	Evidence	Source mix	Categories	Events	Attention	Fetch health