Prime Agent hits 95.5% on reasoning benchmark, beats human experts on code tasks

Prime Intellect's new open-source tool just scored 95.5% on the ARC-AGI-3 benchmark, outpacing the human experts the test was designed around. The startup, valued at $1 billion last month, dropped Prime Agent this week as the first major output from its $130 million Series A round closed in July.

The tool works differently from typical AI coding assistants. Instead of treating each task in isolation, Prime Agent runs inside a persistent Python environment where context sticks around between steps. The model can call tools, spawn sub-agents for specific jobs, and crucially, rewrite its own approach based on what actually works. No static prompts needed.

How the self-improving loop actually functions

Prime Agent sits on what researchers call a Recursive Language Model framework, built around something called a Continual Harness. The innovation here is treating context as a live variable rather than a fixed input. The system can modify its own behavior mid-task without waiting for human feedback or intervention.

Alex Zhang outlined the RLM concept in a blog post back in October 2025, then formalized it in an arXiv paper. Prime Intellect took that theoretical work and turned it into production infrastructure. The underlying model is Opus 5, which handles the actual reasoning and code generation.

Why the benchmark matters

ARC-AGI-3 specifically tests abstract reasoning and novel problem-solving, not just pattern matching on familiar tasks. A 95.5% score beating human-expert baselines means the system solved problems the benchmark creators themselves struggled with. That's different from scoring high on a standard coding test.

The launch follows Prime Intellect's deployment of its verifiers stack version one and work building hundreds of distributed nodes. Investors including Radical Ventures, NVIDIA Ventures, and Intel Capital are watching to see what a $150 million total funding pile can produce in the autonomous AI space.

This article covers product developments and benchmark results. It's informational only and not investment advice.