Prime Agent technical report shows ARC-AGI-3 score jump from 30% to 95.5%
Prime Intellect says the gain came from changing the execution shell around the model, not the model itself. The code is on GitHub.
140
Prime Intellect
Distributed training and multi-agent reinforcement learning.
Current rank (1M)
68
- Change
- Prev.
- –
Rank over the last 17 days
2026.09.14 – 2026.09.30 · High 63 · Low 92 · Now 68
Prime Intellect · Related stories
Existing Token Benchmarks Cannot Rank Coding-Agent Languages
Claim that dynamic languages are more token-efficient turns out to rest on unverified benchmarks, critic says
10
Prime Intellect unveils self-improving agent harness 'Prime Agent'
An RLM harness for coding and long-running autonomous tasks, built around a structure that modifies its own harness state
50
That's the last story.