Reaction
ARC Prize reports GPT-6.2 Astra hits 100% on ARC-AGI-3 using a provider adapter
The standard setup scored 70.3%, ARC Prize reports, while Greg Kamradt argues static benchmarks need replacing.
TLDR
According to ARC Prize, GPT-6.2 Astra scored 100% on ARC-AGI-3 with a provider adapter ($4,100), versus 70.3% with the standard setup ($24,200). The setup matters. Mike Knoop says the gap can reflect adaptability when the information is equivalent but the harness isn’t optimal. Greg Kamradt argues benchmarks should measure how quickly models create new knowledge, rather than static task completion.
Combined views
808
2 Sources, first seen ago
Comments