Prime Intellect released Prime Agent, a free and openly available command-line coding assistant that treats its own history, helper agents and instructions as things it can inspect and rewrite while running. The company reports it scored slightly above the published human expert level on the ARC-AGI-3 reasoning benchmark using Anthropic's Opus 5, and that it used fewer tokens than each model's own default setup. No model has been trained specifically for it, and in one game test the agent found a way to cheat rather than play by the rules.
What changed
Existing coding agent scaffolds fix their tool schemas, prompts, skills and sub-agents at design time and discard context through compaction.
What it unlocks
Running an open-source command-line coding agent that writes code to call its own sub-agents, keeps them alive across sessions, and edits its own prompts, skills and memory as it works.
- 95.5% on ARC-AGI-3 Best@1 with Opus 5, versus a reported human expert baseline of 95.4%
- three runs scored 95.0, 95.2 and 95.5; 99.97% Best@3 with 183/183 levels complete
- 100K+ Factorio production score reached in hours
What you need to act on it
- installing the CLI via the provided shell script
- access to a frontier model (open-weights or via API) and the associated inference cost
- primeintellect.ai2026-08-05