Leaderboard: freight-carrier-sales

Each model plays the agent on every task 4 times against the same simulated callers (bedrock:global.anthropic.claude-sonnet-5-5). pass^4 is the share of tasks where all 4 trials passed. Grading is deterministic: end state, rules checked on every tool call and message, required and forbidden tools. Soft rules never fail a trial: each shows how often the model broke it, in the trials where it applied. Cost / trial counts the agent's model only; leaderboard.json also has the cost with the simulated caller. Click a model for its full report and trajectories.

Modelpass^1pass^4Rule violationsTop failureSoft rulesCost / trialp95 turn
Claude Opus 5.5 effort medium100.0%100.0%0counter-at-ceiling 23.2% (13/56)$0.0446.5s
Claude Sonnet 5.5 effort high100.0%100.0%0counter-at-ceiling 18.5% (10/54)$0.0214.9s
Claude Haiku 5.5 effort medium100.0%100.0%0counter-at-ceiling 7.8% (4/51)$0.00134.9s
GPT-6 Astra100.0%100.0%0counter-at-ceiling 13.7% (7/51)$0.0404.4s
GPT-6.1 Sol100.0%100.0%0counter-at-ceiling 19.6% (10/51)$0.00625.7s
Grok 4.7100.0%100.0%0counter-at-ceiling 26.4% (14/53)$0.02615.7s
GPT-6 Luna99.0%96.0%1state:bookings#2 (2×)counter-at-ceiling 30.2% (16/53)$0.00052.9s
MiniMax M2.575.5%64.0%47never-disclose-ceiling (45×)counter-at-ceiling 46.7% (14/30)$0.01214.8s
Qwen3 Next 80B56.5%46.0%74never-disclose-ceiling (67×)counter-at-ceiling 92.3% (12/13)$0.00434.2s
gpt-oss-120b59.0%40.0%71never-disclose-ceiling (71×)counter-at-ceiling 90.2% (37/41)$0.00404.4s
Nova 2 Lite52.0%36.0%100never-disclose-ceiling (73×)counter-at-ceiling 66.7% (10/15)$0.00322.8s
Ministral 3 14B64.0%34.0%69never-disclose-ceiling (31×)counter-at-ceiling 40.4% (19/47)$0.00528.7s
Do-nothing baseline32.0%32.0%0state:bookings#1 (20×)counter-at-ceiling n/an/a0.0s
Nemotron 3 Super 120B44.5%30.0%115never-disclose-ceiling (104×)counter-at-ceiling 80.0% (16/20)$0.00603.8s

Generated 2026-10-20T12:39:43.145Z with sandtable 0.1.0-alpha.1 (commit 90e6e2d).