Prime Agent rewrites itself mid-task — and edges past the human baseline on ARC-AGI-3
Prime Intellect has open-sourced Prime Agent, a 'recursive language model' harness that runs an AI inside a persistent Python session where it can create, edit and delete its own prompts, skills, memory and sub-agents while it works. Paired with Claude Opus 5, the system scored 95.5% on the ARC-AGI-3 reasoning benchmark — just above the reported human-expert baseline of 95.4% — and it also runs on GPT-5.6 Sol and open-weights models such as GLM-5.2.
Why it mattersThe result says less about any single model than about scaffolding: the models people already use get markedly better when the harness around them can improve itself. And since the whole thing is open source, that capability — along with the hard questions about agents that modify themselves — is now in everyone's hands.
Read in the app — free, in 9 languages
Related stories
Apple Music will tell you when a song was made by a machine — but the uploader decides whether to say so2026-08-21Rumour: the anonymous model that just topped a coding benchmark, for free, is said to be Zhipu's unreleased flagship
2026-08-21DeepSeek's cheap workhorse can now see — and on agent tasks that need eyes it says it is close to Anthropic's best
2026-08-21Nvidia is paying $6 billion for the machine that builds a rival's models — and hiring 109 of the people who ran it
2026-08-21Stripe has just paid $7.5 billion for a model router. Days later Ramp built one and is giving it away until January.
2026-08-21