Every sourced cost-linked effort value for this exact version. Lines connect complete sweeps only.
20
Frontier-Bench v0.1 · agentic terminal coding
Cost / attempt
Anthropic · Anthropic reportClaude Fable 5
≈15.9approx. · 2.7× cost
EffortScoreCost/attempt
max≈33.7%≈$28.00
xhigh≈31.6%≈$20.00
high≈28.8%≈$14.60
medium≈24.8%≈$12.00
low≈17.8%≈$10.30
Anthropic · Anthropic reportClaude Opus 4.8
≈12.3approx. · 3.4× cost
EffortScoreCost/attempt
max≈18.8%≈$16.50
xhigh≈15.5%≈$12.50
high≈12.8%≈$8.00
medium≈9.5%≈$7.30
low≈6.5%≈$4.80
Anthropic · Anthropic reportClaude Opus 5
≈17.7approx. · 3.0× cost
EffortScoreCost/attempt
max≈43.2%≈$16.60
xhigh≈44.3%≈$14.70
high≈39.3%≈$11.60
medium≈34.8%≈$8.30
low≈25.5%≈$5.50
Anthropic · Anthropic reportGPT-5.6 Sol
≈35.5approx. · 10.5× cost
EffortScoreCost/attempt
max≈37.5%≈$11.50
xhigh≈29.0%≈$5.70
high≈22.5%≈$3.60
medium≈14.0%≈$2.60
low≈2.0%≈$1.10
Definition and comparison boundaryinternal methodology
Can the model complete substantial programming work under this benchmark's agent setup?
Software implementation, debugging, or repository work under the published evaluation protocol. Higher is better. The value is the percentage reported in this lab's table.
This is a publisher-defined internal evaluation. The task set or grading details are not fully public, so treat it as directional evidence.
Internal run; mini-SWE-agent harness on GKE, mean reward over five attempts, with Opus 4.8 fallback on safety-classifier refusals for Opus 5 and Fable 5.
