Research
Papers, benchmarks and evals.
-
Replit’s ViBench 2.0 finds no open-weight model can finish a full app build
-
Cognition’s FrontierCode 2.0 finds even the best coding model scores under half
-
SCI-AGENT-EXAM grades AI agents on 120 expert research tasks, and the best clears the bar in about a quarter of runs
-
Salesforce researchers combine eight enterprise agent benchmarks into one EGI-Bench score
-
ETH Zurich’s HPCAgent-Bench finds extra agent tools rarely make code faster but can triple the cost
-
G2 survey finds two-thirds of companies run AI agents in production, with a ‘distrust tax’ on the gains
-
SkillForgeBench finds top AI agents fully finish only 40.9% of professional tasks
-
Long-Horizon-Terminal-Bench 2.0 gives agents three hours and tracks how far they get