Many models.
One effortless connection.

For each metric, we add the reported scores and divide by the number of models with a result. Each model has equal weight. Missing results are excluded, never counted as zero. The displayed range shows the lowest and highest model scores; it is not a confidence interval.

All 11 explorer metrics use the Artificial Analysis snapshot reviewed above, with one configuration per model. Intelligence Index v4.3 is kept separate from earlier versions. Percentages, index points, and Elo ratings are never combined into one overall score. An average Elo is descriptive, not an independently evaluated ensemble rating.

Max/high reasoning configurations are explicitly labeled; these are not confirmed Auto-mode settings. Model availability depends on connected providers. This is a curated review of public evidence, not an exhaustive list of every benchmark on the internet.

Provider-reported results use different agents, budgets, and settings. They are shown separately from the independent averages above. Terminal-Bench 2.1 and 4.0 are different tests.

GPT 5.6 Luna · StateM research submission

The authors report Terminal-Bench 2.1 accuracy increasing from 76.7% to 85.4% with their frozen StateM agent profile. This is a community research result, not an allmodelapi evaluation. It demonstrates how agent software can change a model’s score, so we do not pool it with the independent results.

Community coverage here: 1 of 6 models. We have not independently reproduced these runs. Additional user submissions need an exact model version, benchmark version, configuration, and original results before inclusion.