system-one-bench

~/system-one-bench / compare/celeris-1-decision-vs-perplexity-decider-v1-2

Head to head

Celeris-1 Decision vs Decider V1.2

Celeris-1 Decision scores 83.6 and Decider V1.2 84.6 on the System One Index. Celeris-1 Decision is more accurate on 4 of 16 text datasets, Decider V1.2 on 9, and 3 are tied. Same cases, same request, same scorer.

Overall

Celeris-1 DecisionDecider V1.2
System One Index83.684.6
English83.282.0
Arabic83.186.9
Vision Index–83.8
Calibration error2.6%0.7%
Median latency–178 ms
Cost per 1,000 calls$0.014$0.007

Text datasets, accuracy

Celeris-1 DecisionDecider V1.2
AG News English90.591.5
SST-2 English95.595.5
SST-5 English61.060.0
DAIR Emotion English72.560.0
Banking77 English89.088.0
BoolQ English89.591.5
XNLI English English91.394.0
Typed decisions English76.475.1
XNLI French French90.090.0
XNLI Arabic Arabic80.090.0
MASSIVE scenario Arabic73.579.0
MASSIVE intent Arabic88.091.0
OSACT4 offensive Arabic94.595.0
Tweet sentiment Arabic69.075.5
MIRACL rerank Arabic86.586.5
MIRACL pairwise Arabic90.091.5

Image datasets, accuracy

Celeris-1 DecisionDecider V1.2
SST-2, as an image–94.5
Arabic sentiment, as an image–57.0
Oxford-IIIT Pet–85.5
VQAv2 yes/no–86.5
CountBench–95.5

Model pages: Celeris-1 Decision · Decider V1.2. Latency for API models is the median per call through OpenRouter, timed from a cloud server with 8 calls in flight.