Prime Intellect's Prime Agent lifts ARC-AGI-3 from 30% to 95.5%

Prime Intellect claims its new open-source harness lifts ARC-AGI-3 RHAE Best@1 from 30% to 95.5%, a jump the preprint frames as pushing measurement toward "the model's true maximal underlying capability" rather than a real capability gain.

The paper, posted to arXiv on 24 August by Seth Karten, Alex L. Zhang and colleagues at Prime Intellect, describes Prime Agent as two abstractions bolted together. A persistent IPython REPL follows what the authors call the Recursive Language Model abstraction for programmatic context processing. A separate Continual Harness then "preserves histories, memories, skills, prompts, and subagent specifications across trajectories." The authors argue the design "prevents harness failures from becoming model failures."

Beyond the ARC-AGI-3 headline, the abstract says Prime Agent "matches or exceeds native and popular harnesses across long-context coding, GPU-kernel generation, emulator construction, and autonomous nanoGPT speedruns." On Factorio, "refinement allows for continuous technology progression and dedicated subagents enable parallelized work." Companion figures on Prime Intellect's blog attribute the 95.5% score to Opus 5, with three-run consistency of 95.0, 95.2 and 95.5 and a Best@3 of 99.97% covering 183 of 183 levels. The abstract itself does not name the model behind its headline number.

The code sits on GitHub. Two researchers on our AI Weekly radar shared the paper within a day of its posting.