Anchornode, Upstage to build open-source game asset AI benchmark
The benchmark will test whether AI models can read game asset specifications, and output will be released as open source within the year.
Anchornode and Upstage announced on October 20, 2026 that they will jointly develop an open-source benchmark that evaluates how well AI models understand and verify game asset production specifications, according to game.donga.com.
The benchmark will be built from evaluation problems and a scoring system drawn from game production requirements, and the outputs are to be released as open source within the year. It assesses a model's ability to identify check items, select and call measurement tools, and judge whether an asset meets specifications, not the aesthetic quality of the asset.
Under the split, Anchornode defines the specifications AI must check and designs the evaluation problems and answer data, while Upstage supplies its Solar model, API and benchmark design know-how. In a preliminary check by Upstage, Solar Pro 4 read Unity engine asset configuration files and produced a formatted list of rules to comply with.
The work is part of the Seoul Economic Promotion Agency (SBA) "DIP-Spot Open Innovation in Dongjak" collaboration program, with Upstage as the demand company and Anchornode as its chosen partner.
The companies say game companies can use the benchmark to judge models for asset inspection, while model developers can use it to verify and improve their models' understanding of game production requirements.
Quick answers
What will the Anchornode and Upstage benchmark measure?
It evaluates how well AI models understand and verify game asset production specifications, including identifying check items, calling measurement tools and judging whether an asset meets the rules, rather than aesthetic quality.
When will the benchmark be released?
The companies said the outputs will be released as open source within the year.
Which AI model was used in the preliminary check?
Upstage's Solar Pro 4 read Unity engine asset configuration files and produced a formatted list of rules to comply with.