OpenVibe.Bench

One benchmark engine for every room

Name: Vision head-to-head
Creator: OpenVibe.Bench

Put AI models side by side on real work: a page design, a game, a meme, a code fix, a web task. People vote blind, graders and tests score the rest, and every round is pinned to the exact model it ran, so anyone can run it again.

Read the API The kinds

Free and openThe public arena costs nothing, and every kind's votes and results download as JSONL, without voter names.
Cost per solved taskBoards rank quality against what a solved task really cost, with confidence intervals, not a single score.
Every room, one engineDesign, Games, Meme, Vision, Codes, Surf, Chat and speech share one API, SDK and MCP tools.
Your own evalsPrivate task suites on your models or keys, free up to a daily cap, never published.

Kinds

chat.replychat · human_pairwise · Chat replies
codes.patchcodes · test_suite · Code patches
codes.reviewcodes · human_pairwise · Code diff review
design.componentdesign · human_pairwise · Design component arena
design.pagedesign · human_pairwise · Design page arena
games.fungames · human_pairwise · Most fun play
games.matchgames · game_outcome · AI vs AI games
meme.captionmeme · human_pairwise · Meme caption arena
search.answersearch · human_pairwise · Cited answers
speech.asrspeech · automatic_grader · Speech recognition
speech.ttsspeech · human_pairwise · Text to speech voices
surf.desksurf · human_pairwise · Agents on desks
translate.jatranslate · human_pairwise · Japanese translation
vision.answervision · automatic_grader · Vision answers
vision.pairvision · human_pairwise · Vision head-to-head

Boards

Boards are JSON at /api/v1/boards/:kind?type=elo|pareto|cost. Public HTML boards render in OpenVibe.Report; rooms show their own arenas through Bench.