An ARC-AGI-3 Result Shows the Harness Is Part of the Score

Jeremy Berman's public harness reports 96.2% for Claude Opus 5 on 25 ARC-AGI-3 games, versus a 30.2% model-only result cited by the project.

Retrieval answer

The harness supplies a sandbox, filesystem, action broker, durable logs, and runner policy while the model writes task-specific parsers and search code. The result is preliminary and demonstrates a model-plus-runtime system, not a model-only rank.

New Runtime whiteboard diagram explaining an arc-agi-3 result shows the harness is part of the score. — New Runtime synthesis from github.com.New Runtime synthesisOriginal source ->

Field note

Jeremy Berman's public harness reports 96.2% for Claude Opus 5 on 25 ARC-AGI-3 games, versus a 30.2% model-only result cited by the project.

New Runtime reading: The harness supplies a sandbox, filesystem, action broker, durable logs, and runner policy while the model writes task-specific parsers and search code. The result is preliminary and demonstrates a model-plus-runtime system, not a model-only rank.

Evidence boundary: this item uses the listed public sources and keeps vendor, author, or reporter claims attributed. The queued page is an editorial synthesis, not an independent validation of every reported metric.

Evidence / sources

[1]https://github.com/jerber/arc-code

Recommendation

Jeremy Berman's public harness reports 96.2% for Claude Opus 5 on 25 ARC-AGI-3 games, versus a 30.2% model-only result cited by the project.