Prime Intellect's Prime Agent Smashes AI Benchmark with 95.5% Score

Prime Intellect, a distributed AI infrastructure startup valued at $1 billion after its Series A funding round in July 2026, has unveiled Prime Agent, an open-source coding harness that surpasses human experts on key AI benchmark. The tool uses a Recursive Language Model framework to improve itself through self-learning and modification of its behavior.

Prime Agent's core innovation is the Continual Harness, which treats context as a variable in a persistent Python REPL environment. This allows the model to call tools, spin up sub-agents, and modify its approach based on performance. The framework was built upon foundational research by Alex Zhang, who first outlined the RLM concept in a blog post in October 2025.

The system scored 95.5% on the ARC-AGI-3 benchmark, surpassing established human-expert baselines. Prime Intellect's infrastructure backbone has been building up to this moment, with deployment of its verifiers stack and creation of hundreds of thousands of sandboxed environments through its Environment Hub.