Prime Intellect's open-source Prime Agent tops ARC-AGI-3's human-expert baseline using Opus 5

Prime Intellect on August 5 released Prime Agent, an open-source coding harness the company says scored 95.5 percent on the ARC-AGI-3 benchmark using Anthropic’s Opus 5, Prime Intellect said. That edges past the test’s reported human-expert baseline of 95.4 percent.

Prime Agent is built on what the company calls a Recursive Language Model and a “Continual Harness,” an abstraction that lets the system rewrite its own prompts, skills, memory and sub-agents while a task is running rather than following a fixed script. It is fully open source and installable from the command line.

The headline number is a claim, not a verified result. The 95.5 percent Best@1 figure sits one-tenth of a point above the human baseline, a margin thin enough that run-to-run variance could erase it, and Prime Intellect reported it rather than an independent evaluator. The company also reported a 99.97 percent Best@3 across all 183 ARC-AGI-3 levels and competitive results on long-context tests including OOLONG, LongBenchPro and MazeBench, using fewer tokens than native model harnesses.

None of those figures has been independently reproduced, and “self-improving” here means runtime reconfiguration, not a model that retrains itself. The open-source release does let outside developers test the claims directly.

Prime Intellect closed a $130 million Series A in July, bringing total funding above $150 million. The company is betting that a smarter harness around a frontier model can beat the model’s own tooling, a claim its leaderboard entry now invites others to check.