{"benchmark":{"benchmark_id":"financial-model-benchmark","deterministic_points":"600","sub_benchmark_id":"FMB-FIN-CAPBUD-01","tasks_total":60,"total_points":"600"},"generated_at":"2026-07-23T16:03:02.625046Z","runs":[{"campaign_id":"005d2331-6b72-4c05-991b-fed12aae66f8","completed_at":"2026-07-23T13:03:46.000756Z","execution":{"attempted":60,"invalid_json":0,"missing":0,"provider_failed":0,"schema_failed":2,"tasks_total":60,"valid":58},"model":{"display_name":"nemotron-3-ultra-550b-a55b","provider":"together","reasoning_effort":"high","slug":"nvidia/nemotron-3-ultra-550b-a55b"},"score":{"deterministic":{"earned_points":"60","maximum_points":"600"},"earned_points":"60","maximum_points":"600","percentage":"10.0000","rank_eligible":true},"started_at":"2026-07-23T13:01:53.686623Z","status":"final_deterministic","updated_at":"2026-07-23T13:03:46.000756Z","usage":{"actual_cost_usd":"0.0721842","completion_tokens":10755,"prompt_tokens":55777,"total_tokens":66532},"verification":{"level":"deterministic"}},{"campaign_id":"1a3f7b17-47ee-42d3-889a-78ab32b7c2a1","completed_at":"2026-07-23T13:01:27.232972Z","execution":{"attempted":60,"invalid_json":0,"missing":0,"provider_failed":0,"schema_failed":0,"tasks_total":60,"valid":60},"model":{"display_name":"grok-4.5","provider":"xai","reasoning_effort":"high","slug":"x-ai/grok-4.5"},"score":{"deterministic":{"earned_points":"350","maximum_points":"600"},"earned_points":"350","maximum_points":"600","percentage":"58.3333","rank_eligible":true},"started_at":"2026-07-23T12:53:11.745632Z","status":"final_deterministic","updated_at":"2026-07-23T13:01:27.232972Z","usage":{"actual_cost_usd":"0.766462","completion_tokens":106671,"prompt_tokens":76274,"total_tokens":182945},"verification":{"level":"deterministic"}},{"campaign_id":"2808211c-9430-4158-8a55-fd64c481cd84","completed_at":"2026-07-23T10:03:40.612518Z","execution":{"attempted":60,"invalid_json":0,"missing":0,"provider_failed":0,"schema_failed":0,"tasks_total":60,"valid":60},"model":{"display_name":"glm-5.2","provider":"together","reasoning_effort":"xhigh","slug":"z-ai/glm-5.2"},"score":{"deterministic":{"earned_points":"70","maximum_points":"600"},"earned_points":"70","maximum_points":"600","percentage":"11.6667","rank_eligible":true},"started_at":"2026-07-23T10:01:39.399866Z","status":"final_deterministic","updated_at":"2026-07-23T10:03:40.612518Z","usage":{"actual_cost_usd":"0.09446916","completion_tokens":7151,"prompt_tokens":54853,"total_tokens":62004},"verification":{"level":"deterministic"}},{"campaign_id":"54ce536d-3ee1-44bb-8329-ad1062b0396d","completed_at":"2026-07-23T13:20:48.812869Z","execution":{"attempted":60,"invalid_json":0,"missing":0,"provider_failed":0,"schema_failed":1,"tasks_total":60,"valid":59},"model":{"display_name":"minimax-m3","provider":"morph","reasoning_effort":"max","slug":"minimax/minimax-m3"},"score":{"deterministic":{"earned_points":"60","maximum_points":"600"},"earned_points":"60","maximum_points":"600","percentage":"10.0000","rank_eligible":true},"started_at":"2026-07-23T13:17:40.069864Z","status":"final_deterministic","updated_at":"2026-07-23T13:20:48.812869Z","usage":{"actual_cost_usd":"0.0587094","completion_tokens":8598,"prompt_tokens":63457,"total_tokens":72055},"verification":{"level":"deterministic"}},{"campaign_id":"823eeaa4-afca-435e-8e9b-f512af82f933","completed_at":"2026-07-23T09:45:54.506367Z","execution":{"attempted":60,"invalid_json":13,"missing":0,"provider_failed":0,"schema_failed":0,"tasks_total":60,"valid":47},"model":{"display_name":"gemini-3.5-flash","provider":"google-vertex/global","reasoning_effort":"high","slug":"google/gemini-3.5-flash"},"score":{"deterministic":{"earned_points":"330","maximum_points":"600"},"earned_points":"330","maximum_points":"600","percentage":"55.0000","rank_eligible":true},"started_at":"2026-07-23T09:42:46.962358Z","status":"final_deterministic","updated_at":"2026-07-23T09:45:54.506367Z","usage":{"actual_cost_usd":"1.307316","completion_tokens":133678,"prompt_tokens":69476,"total_tokens":203154},"verification":{"level":"deterministic"}},{"campaign_id":"e1041cd4-56d9-4a92-b0e6-d9176375ca08","completed_at":"2026-07-23T10:10:23.736173Z","execution":{"attempted":60,"invalid_json":0,"missing":0,"provider_failed":0,"schema_failed":0,"tasks_total":60,"valid":60},"model":{"display_name":"kimi-k3","provider":"moonshotai/int4","reasoning_effort":"max","slug":"moonshotai/kimi-k3"},"score":{"deterministic":{"earned_points":"520","maximum_points":"600"},"earned_points":"520","maximum_points":"600","percentage":"86.6667","rank_eligible":true},"started_at":"2026-07-23T08:50:28.124896Z","status":"final_deterministic","updated_at":"2026-07-23T10:10:23.736173Z","usage":{"actual_cost_usd":"1.8896574","completion_tokens":118731,"prompt_tokens":74938,"total_tokens":193669},"verification":{"level":"deterministic"}}],"schema_version":"fmb-live-runs-v3"}
