Report
Top AI reasoning effort reportedly costs up to 42× more on tasks low effort can already solve
A researcher says a high-effort success cannot show whether low effort would work without rerunning the task.
TLDR
A researcher says the team tested models across reasoning-effort levels and benchmarks, finding that top effort costs up to 42× more on tasks low effort can already solve. A successful high-effort run does not show whether low effort would have worked without another run. A collaborator says the team spent $25,000 on the project and open-sourced its trajectories.
Combined views
513
3 Sources, first seen ago
