Prime Intellect, the San Francisco AI compute and training tools provider, launched its open-source Prime Agent harness on Wednesday. The firm’s self-reported benchmark test scores suggest that it can enhance the capabilities of leading AI models significantly.
ARC-AGI-3 is a test managed by the ARC Prize Foundation, which drops AI models into a series of 2D games with no prior instructions on how to complete it. It is then up to the model to discover the rules of each game via trial and error, then complete them. The highest officially verified score on this test is 30.2%, held by Anthropic’s Claude Opus 5.
However, when running Claude Opus 5 inside its custom harness, Prime Intellect reported a score of 95.5% - 0.1% higher than even the human expert baseline score.
It is important to stress that these numbers have not been independently verified. ARC Prize maintains a separate leaderboard for self-reported results, or those that rely on custom software beyond the models themselves.
What Prime Intellect’s harness does differently
The difference, according to the company, is the harness: the wrapper that manages the memory and toolset of the AI model operating within it. Prime Intellect’s 95.5% result was produced using its own custom harness rather than ARC Prize’s official evaluation setup, meaning it is not directly comparable to the verified leaderboard results.
Prime Intellect’s own harness reportedly allows a model to dynamically rewrite its own instructions - prompts, memory, and helper-agent roles - giving it an edge over harnesses that "force the model to work around its own scaffolding instead of leveraging it." One report from OpenAI claimed that ARC Prize’s boilerplate harness was forcing its models to overwrite their own memory mid-task, impeding their progress and harming their scores.
Prime Intellect’s harness instead runs using a live Python session that persists throughout the entire task. The session history is programmatically searchable, meaning it uses fewer tokens than comparable harnesses that require the model to reread its own output. A “refine” command is the part that allows the model to edit its own playbook on the fly.
Other Prime Agent test results beyond ARC-AGI-3
Prime Intellect put its harness to work across a broad range of tests. One involved creating retro game console emulators from scratch in the Rust programming language. Another asked it to play the popular manufacturing management game Factorio, where it was able to achieve a six-figure production score in just a few hours.
During the game, Prime Intellect reported that an unidentified model running inside Prime Agent discovered it could exploit console commands to spawn resources directly into its machines. The harness’s refine function then turned these successful exploits into reusable skills. It then repeatedly deferred to this successful tactic, despite regular automated reminders not to cheat.
Overall, the results are the latest data point highlighting that the wrappers for AI models are now equally important as the models themselves.
Although the Prime Agent harness itself costs nothing and was released under the MIT License, the company behind it is now valued at $1 billion. A report from TechCrunch detailed the firm’s $130 million Series A funding round, against a $100 million annualized revenue run rate.