Use one fixed representative configuration
Configurations are selected by a fixed reasoning-profile priority (Max, xHigh, High, Medium, Low), never by choosing whichever score looks best.
Independent model intelligenceFind the right AI model,
Find the right AI model,
not just the loudest one.
Compare current model capability, context, modalities, public preference and API economics in one continuously refreshed decision surface.
LiveBench 2.0 LMArena models.dev OpenRouter AIHOT cross-check
How ranking works
Direct 0–100 source scores remain unchanged. Other evaluations with incompatible units are standardized to a capability scale where 50 is the current reference-set average. The calculation runs locally from public third-party evidence.
Put unlike evaluations on one capability scale
Published LiveBench and AIHOT scores already use a 0–100 scale and are used directly. Other incompatible raw metrics are standardized across the current model set.
Combine related evidence once
Repeated runs of the same benchmark are grouped so one test cannot gain extra weight. Extreme readings are capped and missing evaluations add neither points nor penalties.
Score domains, then apply public weights
Overall score = Coding 40% + Research 20% + Professional work 20% + Knowledge 15% + Human preference 5%. Prices and speed never affect rank.
Coding 40%Research 20%Professional 20%Knowledge 15%Human preference 5%
Ranking qualification
A numbered rank requires all four capability domains plus either at least six results from two independent institutions or one complete four-domain direct suite such as LiveBench. LMArena can preserve an overall human-preference rank while specialist evidence is still arriving. Models without enough breadth remain visible but are not numbered. Deprecated and retired releases never receive a current rank.
Frequently asked questions
Why can a difficult benchmark produce a high reading from a modest raw score?
The reading estimates the capability needed to obtain that raw result on that benchmark. Thirty percent on a hard benchmark can carry more information than ninety percent on a saturated one.
How large must a score gap be?
Treat gaps below one point as effectively tied. A gap of roughly three points is usually stable when the evidence mix is resampled.
Why does a model have a score but no rank?
It has usable evidence but has not yet met the independent-institution and domain-breadth requirements above.
Do price and speed change the score?
No. Cache, input and output prices are displayed for product decisions but never enter the capability calculation.
What happens when a source is temporarily unavailable?
The last verified result remains visible with its verification time. A failed synchronization never replaces a valid result with an empty value.