ChatGPT
Claude01 / Scope
What Oakgen curates
Oakgen Arena presents a dated OpenArt reference snapshot with original guidance for choosing image and video models. These scores are OpenArt's published results. Oakgen has not conducted or independently reproduced this evaluation.
Index version 1.0 was published on September 15, 2026. We collected the public results on October 30, 2026 UTC.
The collection contains seven video boards and five image boards. Changing a category selects another published board. Sorting a table changes its display order, never the source rank or score.
02 / Source method
How OpenArt ranks models
Blind pairs
Judges compare outputs without model identities, using one named criterion at a time.
Relative scores
A Bradley-Terry model estimates criterion scores relative to a fixed anchor.
Published weights
Board scores combine criteria using the snapshot's weights. Some general fits are reused across boards.
Show uncertainty
The source uses judge-cluster bootstrap estimates for 95% confidence intervals.
03 / Interpretation
A score is a comparison, not a quality grade
Read the score together with its interval and the task being evaluated. A small numerical lead deserves caution when estimates overlap. An ordinal rank alone does not establish a meaningful difference. Compare models within the same board. Different criteria and weights make cross-board score comparisons misleading.
The Anchor label identifies the model used to fix the scale. Its missing interval is labeled explicitly. It does not mean that the model has no uncertainty.
04 / Evidence
Keep the gaps visible
This snapshot does not report composite distinct-judge counts or prices. We leave those values unknown. Vote appearances are not unique voters and must not be added across boards. Reused criterion fits are identified where the source supplies that information.
Raw votes and the complete generation history are not part of this collection. We cannot audit prompt selection, retries or individual judging decisions from aggregate scores.
05 / Your workflow
Build a shortlist. Then test your brief.
Choose a board that matches your task, then compare a few models with the same prompt and reference assets. Record the model version, settings, cost and failures. Inspect the details your project needs, such as text accuracy or subject consistency. Check current availability and licensing separately before production.