Meta researchers taught an 8B AI model to match Claude Opus 4.5 — without the frontier price tag

Summary

Researchers at Meta AI and the University of Illinois Urbana–Champaign introduced EvoHarness-RL, a framework that abstracts the agent's runtime layer and trains the underlying model on when to read, update, or consolidate information from its environment, enabling an 8B-parameter model to match Claude Opus 4.5 on complex enterprise workflows. This research demonstrates that improving the agent harness rather than merely scaling model size can significantly boost AI capabilities, reducing reliance on the high costs of frontier models.