gpt · 2 arms · 6 effortsgpt-5.6-luna

One model across every effort and arm. Each cell is a run — open it for the full card. Failure chips mark runs with no valid slot.

effort strip

effort → output size

Output completion tokens per effort, per arm. A read on how much each model spends as effort rises — not a quality axis.

3.2×low → max output
0100K200Knonelowmediumhighxhighmax103.6K32.1K
codex (OAuth)kiro
gpt-5.6-luna — model · weather-card-benchmark