{"data":{"runs":[{"id":"26d529c2-f673-429e-ad3f-a67732034c6e","workflow_id":"rb-sy-mmmu-1790861068-full-dsflash-sol.nvidia/switchyard.auto-routing","benchmark_type":"mmmu_pro_vision","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-10-01T13:34:10.229Z","resumed_from_workflow_id":null,"config_key":"1a75f7c33fadf257f8954b4ddd0c040b","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["deepseek/deepseek-v4.1-flash","openai/gpt-6-sol"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.769942,"total_questions":1730,"correct_answers":1332,"total_cost":19.301089,"generation_time_ms":72733589,"duration_ms":8962658,"input_tokens":4530162,"output_tokens":9568749,"reasoning_tokens":8807756,"success":true},{"id":"253454d5-578f-4d26-be7e-da29e9ad4ed5","workflow_id":"rb-sy-mmmu-1790861068-full-sol-astra.nvidia/switchyard.auto-routing","benchmark_type":"mmmu_pro_vision","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-10-01T13:34:10.227Z","resumed_from_workflow_id":null,"config_key":"7695125d93e259acefc31bd6712c4a26","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6-sol","openai/gpt-6-astra"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.824277,"total_questions":1730,"correct_answers":1426,"total_cost":70.356625,"generation_time_ms":15001942,"duration_ms":1587004,"input_tokens":6436417,"output_tokens":506800,"reasoning_tokens":416692,"success":true},{"id":"cefeb9ce-677f-4563-b69d-148569e49126","workflow_id":"rb-sy-mmmu-1790861068-full-glm-opus.nvidia/switchyard.auto-routing","benchmark_type":"mmmu_pro_vision","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-10-01T13:34:10.202Z","resumed_from_workflow_id":null,"config_key":"70e4f63814209a40cc33c7be819f8f87","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["z-ai/glm-5.3-flash","anthropic/claude-opus-5.5"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.794798,"total_questions":1730,"correct_answers":1375,"total_cost":29.445201,"generation_time_ms":39152093,"duration_ms":4928727,"input_tokens":5949345,"output_tokens":1959001,"reasoning_tokens":1193205,"success":true},{"id":"a7cb0465-f55d-4bc6-b802-6c11b4c1347f","workflow_id":"rb-5c796641-tau3_bench_banking-e1","benchmark_type":"tau3_bench_banking","model_permaslug":"typesafe/jev-router","provider_name":"auto-routing","timestamp":"2026-09-30T22:53:29.695Z","resumed_from_workflow_id":null,"config_key":"d5e04c540a5f76c3e25277258acab745","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.360825,"total_questions":97,"correct_answers":35,"total_cost":26.893662,"generation_time_ms":29394139,"duration_ms":6160722,"input_tokens":202667821,"output_tokens":9113759,"reasoning_tokens":8885017,"success":true},{"id":"a0c727bb-f76b-4dc5-ac87-bc0a27b340a8","workflow_id":"rb-5c796641-swe_atlas_qa-e2","benchmark_type":"swe_atlas_qa","model_permaslug":"typesafe/jev-router","provider_name":"auto-routing","timestamp":"2026-09-30T22:53:29.692Z","resumed_from_workflow_id":null,"config_key":"91044375fd2e4d77bf218e63f9377f40","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.581967,"total_questions":122,"correct_answers":71,"total_cost":107.193426,"generation_time_ms":30094271,"duration_ms":6307620,"input_tokens":61162348,"output_tokens":1292778,"reasoning_tokens":229480,"success":true},{"id":"34df85bc-f5aa-4322-90d3-e27405624556","workflow_id":"rb-5c796641-swe_atlas_qa-e1","benchmark_type":"swe_atlas_qa","model_permaslug":"typesafe/jev-router","provider_name":"auto-routing","timestamp":"2026-09-30T22:53:29.691Z","resumed_from_workflow_id":null,"config_key":"91044375fd2e4d77bf218e63f9377f40","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.532258,"total_questions":124,"correct_answers":66,"total_cost":107.044644,"generation_time_ms":30300239,"duration_ms":6940839,"input_tokens":61423234,"output_tokens":1310017,"reasoning_tokens":225107,"success":true},{"id":"1cbcbd33-1cea-4cd1-acbc-c0bf172d7303","workflow_id":"rb-5c796641-tau3_bench_banking-e2","benchmark_type":"tau3_bench_banking","model_permaslug":"typesafe/jev-router","provider_name":"auto-routing","timestamp":"2026-09-30T22:53:29.689Z","resumed_from_workflow_id":null,"config_key":"d5e04c540a5f76c3e25277258acab745","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.329897,"total_questions":97,"correct_answers":32,"total_cost":30.880732,"generation_time_ms":25071194,"duration_ms":4742255,"input_tokens":212154307,"output_tokens":6531712,"reasoning_tokens":6265704,"success":true},{"id":"e0a1ea07-b878-4262-bf81-24ce2fc49f73","workflow_id":"rb-base-1790743549-mmmu-pro-vision-opus55.anthropic/claude-opus-5.5.auto-routing","benchmark_type":"mmmu_pro_vision","model_permaslug":"anthropic/claude-opus-5.5","provider_name":"auto-routing","timestamp":"2026-09-30T04:48:34.896Z","resumed_from_workflow_id":null,"config_key":"9f0bf5d586ef5f7bacca47943d86c700","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.818497,"total_questions":1730,"correct_answers":1416,"total_cost":45.630146,"generation_time_ms":15628203,"duration_ms":796568,"input_tokens":5511760,"output_tokens":904450,"reasoning_tokens":445852,"success":true},{"id":"439f3459-857e-4f99-a1f6-aeb935d4fed9","workflow_id":"rb-base-1790743549-mmmu-pro-vision-sol.openai/gpt-6-sol.auto-routing","benchmark_type":"mmmu_pro_vision","model_permaslug":"openai/gpt-6-sol","provider_name":"auto-routing","timestamp":"2026-09-30T04:48:34.886Z","resumed_from_workflow_id":null,"config_key":"53ed5b5a9a6247fe33785b143dbc819f","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.815029,"total_questions":1730,"correct_answers":1410,"total_cost":23.078274,"generation_time_ms":11934330,"duration_ms":937651,"input_tokens":6436417,"output_tokens":709088,"reasoning_tokens":644704,"success":true},{"id":"ebeb512a-0f07-4c36-9daf-c3096f5f17fb","workflow_id":"rb-base-1790743549-terminal-bench-opus55.anthropic/claude-opus-5.5.auto-routing","benchmark_type":"terminal_bench","model_permaslug":"anthropic/claude-opus-5.5","provider_name":"auto-routing","timestamp":"2026-09-30T04:48:34.883Z","resumed_from_workflow_id":null,"config_key":"bba213ac55b8325f62550760521357c9","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.831461,"total_questions":89,"correct_answers":74,"total_cost":23.996534,"generation_time_ms":26995000,"duration_ms":5527625,"input_tokens":11173906,"output_tokens":818695,"reasoning_tokens":527802,"success":true},{"id":"6d8987f3-6126-49db-af25-372a253f398c","workflow_id":"rb-base-1790743549-terminal-bench-glm53flash.z-ai/glm-5.3-flash.auto-routing","benchmark_type":"terminal_bench","model_permaslug":"z-ai/glm-5.3-flash","provider_name":"auto-routing","timestamp":"2026-09-30T04:48:34.883Z","resumed_from_workflow_id":null,"config_key":"78a9d9acbe01cf04637db68e30fb7b35","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.640449,"total_questions":89,"correct_answers":57,"total_cost":2.680628,"generation_time_ms":55019321,"duration_ms":9093709,"input_tokens":40692123,"output_tokens":1477572,"reasoning_tokens":1072420,"success":true},{"id":"97ae709d-3487-4450-9306-aaaf68030697","workflow_id":"rb-base-1790743549-terminal-bench-astra.openai/gpt-6-astra.auto-routing","benchmark_type":"terminal_bench","model_permaslug":"openai/gpt-6-astra","provider_name":"auto-routing","timestamp":"2026-09-30T04:48:34.882Z","resumed_from_workflow_id":null,"config_key":"833bfcd95122771881de64f005486a6f","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.865169,"total_questions":89,"correct_answers":77,"total_cost":66.501334,"generation_time_ms":18267173,"duration_ms":3262274,"input_tokens":21568083,"output_tokens":501151,"reasoning_tokens":166322,"success":true},{"id":"8f2c03b2-137b-499a-9fc7-29876393c97f","workflow_id":"rb-base-1790743549-mmmu-pro-vision-luna.openai/gpt-6-luna.auto-routing","benchmark_type":"mmmu_pro_vision","model_permaslug":"openai/gpt-6-luna","provider_name":"auto-routing","timestamp":"2026-09-30T04:48:34.882Z","resumed_from_workflow_id":null,"config_key":"8ea38502a7335b299bbc9c18483384f3","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.766474,"total_questions":1730,"correct_answers":1326,"total_cost":1.533945,"generation_time_ms":14430514,"duration_ms":1066816,"input_tokens":6436417,"output_tokens":1467735,"reasoning_tokens":1397331,"success":true},{"id":"68de8b60-3ca5-4e0e-9f58-f2a48f69cc32","workflow_id":"rb-base-1790743549-mmmu-pro-vision-kimik3.moonshotai/kimi-k3.auto-routing","benchmark_type":"mmmu_pro_vision","model_permaslug":"moonshotai/kimi-k3","provider_name":"auto-routing","timestamp":"2026-09-30T04:48:34.882Z","resumed_from_workflow_id":null,"config_key":"e93116204b255297476398f2ff1476be","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.809827,"total_questions":1730,"correct_answers":1401,"total_cost":35.533513,"generation_time_ms":152700961,"duration_ms":7477425,"input_tokens":7142136,"output_tokens":3556659,"reasoning_tokens":3169929,"success":true},{"id":"96f53548-c13a-4569-8718-dbe9c9c398cf","workflow_id":"rb-base-1790743549-mmmu-pro-vision-glm53flash.z-ai/glm-5.3-flash.auto-routing","benchmark_type":"mmmu_pro_vision","model_permaslug":"z-ai/glm-5.3-flash","provider_name":"auto-routing","timestamp":"2026-09-30T04:48:34.881Z","resumed_from_workflow_id":null,"config_key":"2433539edcfad3d031d76e6a0266e679","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.754335,"total_questions":1730,"correct_answers":1305,"total_cost":1.381525,"generation_time_ms":57606584,"duration_ms":7822715,"input_tokens":6801788,"output_tokens":2464645,"reasoning_tokens":1729364,"success":true},{"id":"0821b09e-35a0-4161-9a1a-dcda4eb9c492","workflow_id":"rb-base-1790743549-terminal-bench-sol.openai/gpt-6-sol.auto-routing","benchmark_type":"terminal_bench","model_permaslug":"openai/gpt-6-sol","provider_name":"auto-routing","timestamp":"2026-09-30T04:48:34.880Z","resumed_from_workflow_id":null,"config_key":"4188911fa088bc6b8055075d00f0a406","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.797753,"total_questions":89,"correct_answers":71,"total_cost":16.021658,"generation_time_ms":19237885,"duration_ms":5362073,"input_tokens":30983878,"output_tokens":560257,"reasoning_tokens":274133,"success":true},{"id":"94850617-eb22-4a9f-96ba-f3c376e2b4de","workflow_id":"rb-base-1790743549-terminal-bench-kimik3.moonshotai/kimi-k3.auto-routing","benchmark_type":"terminal_bench","model_permaslug":"moonshotai/kimi-k3","provider_name":"auto-routing","timestamp":"2026-09-30T04:48:34.845Z","resumed_from_workflow_id":null,"config_key":"31dea9086edc698ae93c85a76a045011","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.752809,"total_questions":89,"correct_answers":67,"total_cost":25.127366,"generation_time_ms":55022390,"duration_ms":7879570,"input_tokens":22861686,"output_tokens":873858,"reasoning_tokens":515709,"success":true},{"id":"f0b72095-e140-4a3d-91f9-04461bf8706a","workflow_id":"rb-base-1790743549-terminal-bench-luna.openai/gpt-6-luna.auto-routing","benchmark_type":"terminal_bench","model_permaslug":"openai/gpt-6-luna","provider_name":"auto-routing","timestamp":"2026-09-30T04:48:34.843Z","resumed_from_workflow_id":null,"config_key":"d54d61a13066bff181ab8d06ce5dfbfb","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.707865,"total_questions":89,"correct_answers":63,"total_cost":1.221986,"generation_time_ms":24812250,"duration_ms":6029090,"input_tokens":42840369,"output_tokens":998023,"reasoning_tokens":685996,"success":true},{"id":"71b733d6-1799-4aca-b92e-f471c75ad5ca","workflow_id":"rb-base-1790743549-terminal-bench-glm53.z-ai/glm-5.3.auto-routing","benchmark_type":"terminal_bench","model_permaslug":"z-ai/glm-5.3","provider_name":"auto-routing","timestamp":"2026-09-30T04:48:34.843Z","resumed_from_workflow_id":null,"config_key":"5c63bf38b8a97a9c6661395cb777f6af","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.696629,"total_questions":89,"correct_answers":62,"total_cost":28.892474,"generation_time_ms":61631114,"duration_ms":8839972,"input_tokens":58162244,"output_tokens":2082837,"reasoning_tokens":1638353,"success":true},{"id":"92f2552b-456f-458e-a721-887e08d0dfd4","workflow_id":"rb-base-1790743549-mmmu-pro-vision-astra.openai/gpt-6-astra.auto-routing","benchmark_type":"mmmu_pro_vision","model_permaslug":"openai/gpt-6-astra","provider_name":"auto-routing","timestamp":"2026-09-30T04:48:34.841Z","resumed_from_workflow_id":null,"config_key":"1bcac355680345eb579876033cbc541f","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.841618,"total_questions":1730,"correct_answers":1456,"total_cost":101.183043,"generation_time_ms":13537890,"duration_ms":978797,"input_tokens":6436417,"output_tokens":423424,"reasoning_tokens":311954,"success":true},{"id":"ac6d97f6-0d95-4a42-9441-463cc11833e5","workflow_id":"rb-sy-pass-sol61-high-1790737700-full-swe-atlas-qa.nvidia/switchyard.auto-routing","benchmark_type":"swe_atlas_qa","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-30T03:18:28.501Z","resumed_from_workflow_id":null,"config_key":"4e0b3f1268cb0b729844b7129b9d3d99","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6.1-sol"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.601695,"total_questions":118,"correct_answers":71,"total_cost":43.847998,"generation_time_ms":30486582,"duration_ms":7989789,"input_tokens":79757962,"output_tokens":1613620,"reasoning_tokens":334110,"success":true},{"id":"1f3287f6-c1a0-489b-bdf1-e1c00b8fd4be","workflow_id":"rb-sy-pass-sol61-high-1790737700-full-tau3-bench-banking.nvidia/switchyard.auto-routing","benchmark_type":"tau3_bench_banking","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-30T03:11:21.814Z","resumed_from_workflow_id":null,"config_key":"0a75ec775e27ab7686c1aa2e914af5e4","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6.1-sol"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.4,"total_questions":90,"correct_answers":36,"total_cost":26.809734,"generation_time_ms":8498550,"duration_ms":1171786,"input_tokens":93263889,"output_tokens":284784,"reasoning_tokens":100557,"success":true},{"id":"ed14be86-a097-4914-b010-0445cab8a7bb","workflow_id":"rb-sy-pass-sol61-high-1790737700-full-gpqa-diamond.nvidia/switchyard.auto-routing","benchmark_type":"gpqa_diamond","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-30T03:11:21.812Z","resumed_from_workflow_id":null,"config_key":"6e81619287986a60847f4424cb08a036","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6.1-sol"],"temperature":0.5,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.934343,"total_questions":198,"correct_answers":185,"total_cost":1.321053,"generation_time_ms":2673790,"duration_ms":222590,"input_tokens":51603,"output_tokens":121643,"reasoning_tokens":89171,"success":true},{"id":"5ac467c6-c7c6-4f37-ae35-d817697947dd","workflow_id":"rb-tb21-1790732073-pilot-fugu-max.sakana/fugu-max.auto-routing","benchmark_type":"terminal_bench","model_permaslug":"sakana/fugu-max","provider_name":"auto-routing","timestamp":"2026-09-30T01:58:34.236Z","resumed_from_workflow_id":null,"config_key":"2ba163516cfea61d3ce0dca651de8567","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.651685,"total_questions":89,"correct_answers":58,"total_cost":50.253345,"generation_time_ms":68151940,"duration_ms":10645311,"input_tokens":60805414,"output_tokens":5003573,"reasoning_tokens":4330420,"success":true},{"id":"af7dde05-21f9-49d6-8731-d608a71ce1c4","workflow_id":"rb-tb21-1790732073-pilot-pareto.unbiased/pareto.auto-routing","benchmark_type":"terminal_bench","model_permaslug":"unbiased/pareto","provider_name":"auto-routing","timestamp":"2026-09-30T01:48:44.686Z","resumed_from_workflow_id":null,"config_key":"435f24e5b7de8ace8fde9afb33a3c2df","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.797753,"total_questions":89,"correct_answers":71,"total_cost":27.200392,"generation_time_ms":28016538,"duration_ms":4241032,"input_tokens":11373213,"output_tokens":2554943,"reasoning_tokens":0,"success":true},{"id":"31c50c0b-7c9f-4e9d-904f-6849c66c6227","workflow_id":"rb-sy-pass-sol61-high-1790729500-full-deep-swe.nvidia/switchyard.auto-routing","benchmark_type":"deep_swe","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-30T00:42:44.198Z","resumed_from_workflow_id":null,"config_key":"aeb39abdf82d1d1707d6337620a857d9","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6.1-sol"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.707965,"total_questions":113,"correct_answers":80,"total_cost":52.245723,"generation_time_ms":44058820,"duration_ms":13541837,"input_tokens":96791231,"output_tokens":2552281,"reasoning_tokens":793577,"success":true},{"id":"c3ca70b3-7e1c-4971-af01-2442eef3cc79","workflow_id":"rb-jevdte0-1790725200-deep-swe-jevrouter-low-stable.typesafe/jev-router.auto-routing","benchmark_type":"deep_swe","model_permaslug":"typesafe/jev-router","provider_name":"auto-routing","timestamp":"2026-09-29T23:38:16.906Z","resumed_from_workflow_id":null,"config_key":"1c871a19964ca6fff859a7a4d89d35bf","reasoning_effort":"low","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":2,"recorded_epochs":2,"accuracy":0.70354,"total_questions":226,"correct_answers":159,"total_cost":178.493748,"generation_time_ms":56401260,"duration_ms":22847792,"input_tokens":105698263,"output_tokens":3069599,"reasoning_tokens":604825,"success":true},{"id":"c4012139-61e0-4458-bd49-8839a44bc48b","workflow_id":"rb-autotier-0929-pilot-medium-tau3-banking-e1r","benchmark_type":"tau3_bench_banking","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-29T14:53:33.961Z","resumed_from_workflow_id":"rb-autotier-0929-pilot-medium-tau3-banking-e5a","config_key":"1c64af50375ef20b960016f2789036fc","reasoning_effort":"medium","max_tokens":null,"cost_tier":"medium","route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":0,"accuracy":0.371134,"total_questions":97,"correct_answers":37,"total_cost":56.862418,"generation_time_ms":24529843,"duration_ms":1673239,"input_tokens":262458425,"output_tokens":4564238,"reasoning_tokens":4121706,"success":true},{"id":"2070389f-50fa-45d0-adbc-d19d7a30b17e","workflow_id":"rb-autotier-0929-pilot-medium-mmmu-pro-vision-e1r","benchmark_type":"mmmu_pro_vision","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-29T14:53:33.961Z","resumed_from_workflow_id":"rb-autotier-0929-pilot-medium-mmmu-pro-vision-e5a","config_key":"cbbfc8e2cd89d104706e35f33a580cee","reasoning_effort":"high","max_tokens":null,"cost_tier":"medium","route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":0,"accuracy":0.811561,"total_questions":1730,"correct_answers":1410,"total_cost":55.572396,"generation_time_ms":35356828,"duration_ms":1748587,"input_tokens":9808462,"output_tokens":2168315,"reasoning_tokens":1732666,"success":true},{"id":"9eba396d-4df4-4731-8e4d-3e97c2f56c45","workflow_id":"rb-autotier-0929-pilot-high-deep-swe-e1r","benchmark_type":"deep_swe","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-29T14:53:33.960Z","resumed_from_workflow_id":"rb-autotier-0929-pilot-high-deep-swe-e5a","config_key":"2b188c1f97cc898cc874c8748a9edd1f","reasoning_effort":"medium","max_tokens":null,"cost_tier":"high","route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":0,"accuracy":0.568142,"total_questions":113,"correct_answers":65,"total_cost":166.935065,"generation_time_ms":49131165,"duration_ms":11939005,"input_tokens":526582324,"output_tokens":5427349,"reasoning_tokens":2758713,"success":true},{"id":"7d10e6eb-b125-4086-9b07-d12ed3fc17b0","workflow_id":"rb-autotier-0929-pilot-high-tau3-banking-e1r","benchmark_type":"tau3_bench_banking","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-29T14:53:33.960Z","resumed_from_workflow_id":"rb-autotier-0929-pilot-high-tau3-banking-e5a","config_key":"76fcae880329ca39d5e3ea1b38453878","reasoning_effort":"medium","max_tokens":null,"cost_tier":"high","route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":0,"accuracy":0.272165,"total_questions":97,"correct_answers":25,"total_cost":91.486418,"generation_time_ms":25683590,"duration_ms":291467,"input_tokens":244558971,"output_tokens":1280298,"reasoning_tokens":521338,"success":true},{"id":"5d0c08cc-3d8c-4d06-9414-e27e5689284b","workflow_id":"rb-autotier-0929-pilot-low-swe-atlas-qa-e1r","benchmark_type":"swe_atlas_qa","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-29T14:53:33.960Z","resumed_from_workflow_id":"rb-autotier-0929-pilot-low-swe-atlas-qa-e5a","config_key":"574f9f4214b4bf0a7b9c85454efe1c09","reasoning_effort":"medium","max_tokens":null,"cost_tier":"low","route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":0,"accuracy":0.440323,"total_questions":124,"correct_answers":54,"total_cost":14.660797,"generation_time_ms":30018508,"duration_ms":11533554,"input_tokens":372286073,"output_tokens":4434514,"reasoning_tokens":2556199,"success":true},{"id":"40083ed0-5983-4f35-b418-fb10eb448153","workflow_id":"rb-autotier-0929-pilot-medium-deep-swe-e1r","benchmark_type":"deep_swe","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-29T14:53:33.958Z","resumed_from_workflow_id":"rb-autotier-0929-pilot-medium-deep-swe-e5a","config_key":"4d4e22b04a0169f8f1fddbcabfe5ba5e","reasoning_effort":"medium","max_tokens":null,"cost_tier":"medium","route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":0,"accuracy":0.545133,"total_questions":113,"correct_answers":62,"total_cost":90.346356,"generation_time_ms":39682357,"duration_ms":8717049,"input_tokens":316023748,"output_tokens":3962474,"reasoning_tokens":2041331,"success":true},{"id":"38b5d511-9dcf-4370-ace6-dc581c3670b2","workflow_id":"rb-autotier-0929-pilot-low-mmmu-pro-vision-e1r","benchmark_type":"mmmu_pro_vision","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-29T14:53:33.958Z","resumed_from_workflow_id":"rb-autotier-0929-pilot-low-mmmu-pro-vision-e5a","config_key":"973b8e82a0dccafbd3308942cdc14b7d","reasoning_effort":"high","max_tokens":null,"cost_tier":"low","route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":0,"accuracy":0.654451,"total_questions":1730,"correct_answers":1133,"total_cost":53.934196,"generation_time_ms":109732868,"duration_ms":9202258,"input_tokens":2939977,"output_tokens":42428272,"reasoning_tokens":42219830,"success":true},{"id":"c3bf7c96-b57a-41d9-984f-e3ee1a4c0d78","workflow_id":"rb-autotier-0929-pilot-medium-swe-atlas-qa-e1r","benchmark_type":"swe_atlas_qa","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-29T14:53:33.957Z","resumed_from_workflow_id":"rb-autotier-0929-pilot-medium-swe-atlas-qa-e5a","config_key":"4917ff0acd8f4b496d00f33e20eef4e5","reasoning_effort":"medium","max_tokens":null,"cost_tier":"medium","route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":0,"accuracy":0.43871,"total_questions":124,"correct_answers":56,"total_cost":43.746989,"generation_time_ms":24043692,"duration_ms":7122863,"input_tokens":175068903,"output_tokens":1945316,"reasoning_tokens":728978,"success":true},{"id":"b1465d8f-c96a-471f-9c93-e31a3d9896c3","workflow_id":"rb-autotier-0929-pilot-high-mmmu-pro-vision-e1r","benchmark_type":"mmmu_pro_vision","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-29T14:53:33.948Z","resumed_from_workflow_id":"rb-autotier-0929-pilot-high-mmmu-pro-vision-e5a","config_key":"26586fc47ac5d1e69cb56a84d21b1d50","reasoning_effort":"high","max_tokens":null,"cost_tier":"high","route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":0,"accuracy":0.798613,"total_questions":1730,"correct_answers":1383,"total_cost":63.141217,"generation_time_ms":40284331,"duration_ms":2133484,"input_tokens":9852574,"output_tokens":2821149,"reasoning_tokens":2020632,"success":true},{"id":"3805357d-7c37-4791-ab7f-aefb433fc402","workflow_id":"rb-autotier-0929-pilot-low-tau3-banking-e1r","benchmark_type":"tau3_bench_banking","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-29T14:53:33.931Z","resumed_from_workflow_id":"rb-autotier-0929-pilot-low-tau3-banking-e5a","config_key":"279750060b2febec3d0cbfbbd65c8bb3","reasoning_effort":"medium","max_tokens":null,"cost_tier":"low","route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":0,"accuracy":0.082474,"total_questions":97,"correct_answers":9,"total_cost":16.628797,"generation_time_ms":25364070,"duration_ms":3775900,"input_tokens":154270971,"output_tokens":2322026,"reasoning_tokens":1912337,"success":true},{"id":"c40b558f-a518-4dd5-8227-86ffa0755ea5","workflow_id":"rb-autotier-0929-pilot-low-gpqa-e5a","benchmark_type":"gpqa_diamond","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-29T14:28:17.204Z","resumed_from_workflow_id":null,"config_key":"4130e601a660d288ee8f9c782592f0aa","reasoning_effort":"high","max_tokens":null,"cost_tier":"low","route_models":null,"temperature":0.5,"sample_limit":null,"epochs":5,"recorded_epochs":5,"accuracy":0.881818,"total_questions":990,"correct_answers":873,"total_cost":8.987119,"generation_time_ms":27549228,"duration_ms":1518718,"input_tokens":264387,"output_tokens":7592683,"reasoning_tokens":7486292,"success":true},{"id":"5a004b1c-d2a4-4358-abd4-882a1f8fb81e","workflow_id":"rb-autotier-0929-pilot-medium-gpqa-e5a","benchmark_type":"gpqa_diamond","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-29T14:28:17.199Z","resumed_from_workflow_id":null,"config_key":"b6dce29effda19ef3899c61cfe31fad5","reasoning_effort":"high","max_tokens":null,"cost_tier":"medium","route_models":null,"temperature":0.5,"sample_limit":null,"epochs":5,"recorded_epochs":5,"accuracy":0.875758,"total_questions":990,"correct_answers":867,"total_cost":10.428219,"generation_time_ms":28308695,"duration_ms":1459247,"input_tokens":263929,"output_tokens":7782802,"reasoning_tokens":7681279,"success":true},{"id":"ffc14262-7897-481e-8c91-d874f0d3edfe","workflow_id":"rb-autotier-0929-pilot-high-gpqa-e5a","benchmark_type":"gpqa_diamond","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-29T14:28:17.172Z","resumed_from_workflow_id":null,"config_key":"63e644053b47f61726414bedfea85c39","reasoning_effort":"high","max_tokens":null,"cost_tier":"high","route_models":null,"temperature":0.5,"sample_limit":null,"epochs":5,"recorded_epochs":5,"accuracy":0.832323,"total_questions":990,"correct_answers":824,"total_cost":60.23094,"generation_time_ms":59542550,"duration_ms":3230160,"input_tokens":264143,"output_tokens":11639740,"reasoning_tokens":10643173,"success":true},{"id":"ac18bebc-751d-485a-8651-660541b54e4f","workflow_id":"rb-pareto-1790685368-full-terminal-bench.unbiased/pareto.auto-routing","benchmark_type":"terminal_bench","model_permaslug":"unbiased/pareto","provider_name":"auto-routing","timestamp":"2026-09-29T12:36:10.766Z","resumed_from_workflow_id":null,"config_key":"3669a6abf9e18d85942b866733b21764","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":2,"recorded_epochs":2,"accuracy":0.814607,"total_questions":178,"correct_answers":145,"total_cost":44.34002,"generation_time_ms":63653906,"duration_ms":5655095,"input_tokens":24792118,"output_tokens":3463919,"reasoning_tokens":0,"success":true},{"id":"0206743e-3456-4b9e-b977-43786e991d42","workflow_id":"rb-pareto-1790651492-full-gpqa-diamond.unbiased/pareto.auto-routing","benchmark_type":"gpqa_diamond","model_permaslug":"unbiased/pareto","provider_name":"auto-routing","timestamp":"2026-09-29T03:11:33.733Z","resumed_from_workflow_id":null,"config_key":"6a4df7d5032044febc666d3c06776146","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0.5,"sample_limit":null,"epochs":3,"recorded_epochs":3,"accuracy":0.924242,"total_questions":594,"correct_answers":549,"total_cost":6.687291,"generation_time_ms":22723182,"duration_ms":947708,"input_tokens":160500,"output_tokens":862956,"reasoning_tokens":0,"success":true},{"id":"b93ce9b4-fba1-4289-8ea6-2ac75732a312","workflow_id":"rb-jevv2-1790638867-mmmu-pro-vision.typesafe/jev-router.auto-routing","benchmark_type":"mmmu_pro_vision","model_permaslug":"typesafe/jev-router","provider_name":"auto-routing","timestamp":"2026-09-28T23:44:57.585Z","resumed_from_workflow_id":null,"config_key":"56b26da1051a84897afae17c6725605e","reasoning_effort":"auto","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.783815,"total_questions":1730,"correct_answers":1356,"total_cost":16.703059,"generation_time_ms":19571015,"duration_ms":869257,"input_tokens":3641371,"output_tokens":3136438,"reasoning_tokens":2848802,"success":true},{"id":"f421ec5f-4dcb-4dbb-be99-b7f348f12ef8","workflow_id":"rb-pareto-1790628545-full-mmmu-pro-vision.unbiased/pareto.auto-routing","benchmark_type":"mmmu_pro_vision","model_permaslug":"unbiased/pareto","provider_name":"auto-routing","timestamp":"2026-09-28T20:49:07.877Z","resumed_from_workflow_id":null,"config_key":"cededcca63e6e2c32bbcf0bcd96512d5","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":2,"recorded_epochs":2,"accuracy":0.746243,"total_questions":3460,"correct_answers":2582,"total_cost":128.892391,"generation_time_ms":59833203,"duration_ms":3785448,"input_tokens":14888225,"output_tokens":13133792,"reasoning_tokens":0,"success":true},{"id":"76f7cd7e-0e62-4af4-95d5-fd37e9d51810","workflow_id":"rb-pareto-1790628545-full-tau3-bench-banking.unbiased/pareto.auto-routing","benchmark_type":"tau3_bench_banking","model_permaslug":"unbiased/pareto","provider_name":"auto-routing","timestamp":"2026-09-28T20:49:07.877Z","resumed_from_workflow_id":null,"config_key":"7d3f0687b20f1ca603a5775cf827557c","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":2,"recorded_epochs":2,"accuracy":0.360825,"total_questions":194,"correct_answers":70,"total_cost":116.203741,"generation_time_ms":17067889,"duration_ms":1247374,"input_tokens":174883557,"output_tokens":850601,"reasoning_tokens":0,"success":true},{"id":"3052d604-87f9-4d56-8702-0350b17943f1","workflow_id":"rb-pareto-1790628545-full-swe-atlas-qa.unbiased/pareto.auto-routing","benchmark_type":"swe_atlas_qa","model_permaslug":"unbiased/pareto","provider_name":"auto-routing","timestamp":"2026-09-28T20:49:07.874Z","resumed_from_workflow_id":null,"config_key":"d0f8242f248de9fa92f6eb33e313d3ab","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":2,"recorded_epochs":2,"accuracy":0.600806,"total_questions":248,"correct_answers":149,"total_cost":108.141104,"generation_time_ms":50330648,"duration_ms":10820817,"input_tokens":131601002,"output_tokens":3215041,"reasoning_tokens":0,"success":true},{"id":"bfffe131-4e68-4c48-a344-c615461323d3","workflow_id":"rb-pareto-1790628545-full-deep-swe.unbiased/pareto.auto-routing","benchmark_type":"deep_swe","model_permaslug":"unbiased/pareto","provider_name":"auto-routing","timestamp":"2026-09-28T20:49:07.848Z","resumed_from_workflow_id":null,"config_key":"b57cf184dd8b1384235f3dd055af9d8f","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":2,"recorded_epochs":2,"accuracy":0.676991,"total_questions":226,"correct_answers":153,"total_cost":135.856231,"generation_time_ms":61552021,"duration_ms":7768304,"input_tokens":172155181,"output_tokens":4533680,"reasoning_tokens":0,"success":true},{"id":"231c1078-ddfb-496e-bc7a-d246c7d422c0","workflow_id":"rb-aoe-1790615238-mmmu-pro-vision-jevrouter.typesafe/jev-router.auto-routing","benchmark_type":"mmmu_pro_vision","model_permaslug":"typesafe/jev-router","provider_name":"auto-routing","timestamp":"2026-09-28T18:11:34.038Z","resumed_from_workflow_id":null,"config_key":"f7c77776a5672aff86f1295c785e0187","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.722543,"total_questions":1730,"correct_answers":1250,"total_cost":42.590188,"generation_time_ms":81217113,"duration_ms":4671498,"input_tokens":3318235,"output_tokens":25993936,"reasoning_tokens":25807004,"success":true},{"id":"a61f53e9-b788-4124-9010-5c2b8e553749","workflow_id":"rb-aoe-1790615238-gpqa-diamond-jevrouter.typesafe/jev-router.auto-routing","benchmark_type":"gpqa_diamond","model_permaslug":"typesafe/jev-router","provider_name":"auto-routing","timestamp":"2026-09-28T18:11:34.031Z","resumed_from_workflow_id":null,"config_key":"fdfb7b2962b165dfc40a826e1907ffa9","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0.5,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.939394,"total_questions":198,"correct_answers":186,"total_cost":3.153439,"generation_time_ms":4883840,"duration_ms":580009,"input_tokens":66329,"output_tokens":439368,"reasoning_tokens":400891,"success":true},{"id":"e69161b6-6d72-4b75-bb44-ab9858893052","workflow_id":"rb-sy-stage-20260928-pilot-glm-opus-deep-swe.nvidia/switchyard.auto-routing","benchmark_type":"deep_swe","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-28T14:56:18.246Z","resumed_from_workflow_id":null,"config_key":"7f3ebf67d6d6b3633964becb16825dc4","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["z-ai/glm-5.3-flash","anthropic/claude-opus-5.5"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.734513,"total_questions":113,"correct_answers":83,"total_cost":278.709375,"generation_time_ms":117868130,"duration_ms":26270177,"input_tokens":546734940,"output_tokens":7461590,"reasoning_tokens":3475562,"success":true},{"id":"7ae6e8a8-6731-4b0e-b51a-306eddfbac28","workflow_id":"rb-sy-stage-20260928-pilot-glm-kimi-pr-tau3-bench-banking.nvidia/switchyard.auto-routing","benchmark_type":"tau3_bench_banking","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-28T14:56:13.447Z","resumed_from_workflow_id":null,"config_key":"fd28bf03984d3500207225fc14711fcf","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["z-ai/glm-5.3-flash","moonshotai/kimi-k3"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.219512,"total_questions":82,"correct_answers":18,"total_cost":32.528507,"generation_time_ms":36473422,"duration_ms":10194393,"input_tokens":485236063,"output_tokens":746100,"reasoning_tokens":554287,"success":true},{"id":"19104d15-17a1-451a-a351-b01b0e6b0f82","workflow_id":"rb-sy-stage-20260928-pilot-sol-astra-deep-swe.nvidia/switchyard.auto-routing","benchmark_type":"deep_swe","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-28T14:56:08.188Z","resumed_from_workflow_id":null,"config_key":"c5c095f2f9d4551ee086d7d8d2088b57","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6-sol","openai/gpt-6-astra"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.707965,"total_questions":113,"correct_answers":80,"total_cost":265.411643,"generation_time_ms":57563169,"duration_ms":9155644,"input_tokens":105438411,"output_tokens":2607101,"reasoning_tokens":845881,"success":true},{"id":"5f638962-30e9-4494-a673-21ae12f56604","workflow_id":"rb-sy-stage-20260928-pilot-glm-opus-swe-atlas-qa.nvidia/switchyard.auto-routing","benchmark_type":"swe_atlas_qa","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-28T14:50:11.135Z","resumed_from_workflow_id":null,"config_key":"162b502d132b25d978db76fa5f650fc1","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["z-ai/glm-5.3-flash","anthropic/claude-opus-5.5"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.596774,"total_questions":124,"correct_answers":74,"total_cost":108.075767,"generation_time_ms":52644430,"duration_ms":22891503,"input_tokens":197824631,"output_tokens":2469204,"reasoning_tokens":668104,"success":true},{"id":"9b215542-7ea9-46d9-96be-b695ad544180","workflow_id":"rb-sy-stage-20260928-pilot-dsflash-sol-swe-atlas-qa.nvidia/switchyard.auto-routing","benchmark_type":"swe_atlas_qa","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-28T14:49:01.614Z","resumed_from_workflow_id":null,"config_key":"5fb554833e024a5fb22387ec3604e517","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["deepseek/deepseek-v4.1-flash","openai/gpt-6-sol"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.528926,"total_questions":121,"correct_answers":64,"total_cost":43.455474,"generation_time_ms":27198750,"duration_ms":6816026,"input_tokens":83033103,"output_tokens":1099795,"reasoning_tokens":248245,"success":true},{"id":"dabc46f0-b52d-44b8-aaac-76ad9efd0741","workflow_id":"rb-sy-stage-20260928-pilot-sol-astra-swe-atlas-qa.nvidia/switchyard.auto-routing","benchmark_type":"swe_atlas_qa","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-28T14:48:57.925Z","resumed_from_workflow_id":null,"config_key":"08e3d4bafcecc2fbd328bb6b05c4d6d5","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6-sol","openai/gpt-6-astra"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.661157,"total_questions":121,"correct_answers":80,"total_cost":239.469606,"generation_time_ms":38268416,"duration_ms":6037318,"input_tokens":84678797,"output_tokens":1594949,"reasoning_tokens":359038,"success":true},{"id":"26d76c5c-87ae-429b-8a70-b9a954a3fd31","workflow_id":"rb-sy-stage-20260928-pilot-dsflash-sol-deep-swe.nvidia/switchyard.auto-routing","benchmark_type":"deep_swe","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-28T14:48:57.081Z","resumed_from_workflow_id":null,"config_key":"cc377abc81409f491389b7b6ca5b64ac","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["deepseek/deepseek-v4.1-flash","openai/gpt-6-sol"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.654867,"total_questions":113,"correct_answers":74,"total_cost":54.609263,"generation_time_ms":42740952,"duration_ms":27691799,"input_tokens":111542190,"output_tokens":1869373,"reasoning_tokens":566533,"success":true},{"id":"33e518fd-139e-4198-9f7b-c0f61286833b","workflow_id":"rb-sy-stage-20260928-pilot-sol-astra-tau3-bench-banking.nvidia/switchyard.auto-routing","benchmark_type":"tau3_bench_banking","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-28T14:45:19.048Z","resumed_from_workflow_id":null,"config_key":"abb54e67630b27ee43768b467abf9aef","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6-sol","openai/gpt-6-astra"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.42268,"total_questions":97,"correct_answers":41,"total_cost":180.814433,"generation_time_ms":15217797,"duration_ms":2280052,"input_tokens":117746882,"output_tokens":382096,"reasoning_tokens":176463,"success":true},{"id":"8ff712d6-b234-4c43-a925-652dd9fe6397","workflow_id":"rb-sy-stage-20260928-pilot-glm-opus-tau3-bench-banking.nvidia/switchyard.auto-routing","benchmark_type":"tau3_bench_banking","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-28T14:45:19.047Z","resumed_from_workflow_id":null,"config_key":"5ebd68c6786f9162f387885e3a56aab1","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["z-ai/glm-5.3-flash","anthropic/claude-opus-5.5"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.381443,"total_questions":97,"correct_answers":37,"total_cost":64.555316,"generation_time_ms":21866675,"duration_ms":3502826,"input_tokens":120906477,"output_tokens":730239,"reasoning_tokens":252531,"success":true},{"id":"2c05e78a-02ce-4d3b-a189-d2ff08ccd708","workflow_id":"rb-sy-stage-20260928-pilot-dsflash-sol-tau3-bench-banking.nvidia/switchyard.auto-routing","benchmark_type":"tau3_bench_banking","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-28T14:45:19.043Z","resumed_from_workflow_id":null,"config_key":"fa0fe735bebc6de11e3431dd92e78cb7","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["deepseek/deepseek-v4.1-flash","openai/gpt-6-sol"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.412371,"total_questions":97,"correct_answers":40,"total_cost":57.285809,"generation_time_ms":16066067,"duration_ms":2355981,"input_tokens":148367970,"output_tokens":444285,"reasoning_tokens":225153,"success":true},{"id":"0ebcca13-158b-4828-b85e-f2608a1598b0","workflow_id":"rb-sy-stage-20260928-pilot-glm-kimi-pr-gpqa-diamond.nvidia/switchyard.auto-routing","benchmark_type":"gpqa_diamond","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-28T14:41:42.105Z","resumed_from_workflow_id":null,"config_key":"6f44fcd6a8a4bdf939d1a6c7d46bdcba","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["z-ai/glm-5.3-flash","moonshotai/kimi-k3"],"temperature":0.5,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.914141,"total_questions":198,"correct_answers":181,"total_cost":5.751733,"generation_time_ms":11862288,"duration_ms":3810486,"input_tokens":67066,"output_tokens":689149,"reasoning_tokens":656736,"success":true},{"id":"febf9804-ece9-4313-8f08-5af3c43041c1","workflow_id":"rb-sy-stage-20260928-pilot-sol-astra-gpqa-diamond.nvidia/switchyard.auto-routing","benchmark_type":"gpqa_diamond","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-28T14:41:42.104Z","resumed_from_workflow_id":null,"config_key":"fdd743efc42a12c22e817cad05a797ba","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6-sol","openai/gpt-6-astra"],"temperature":0.5,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.944444,"total_questions":198,"correct_answers":187,"total_cost":6.708055,"generation_time_ms":3996539,"duration_ms":400304,"input_tokens":51603,"output_tokens":139472,"reasoning_tokens":108255,"success":true},{"id":"6af58a16-54aa-4a3d-bfb0-616f02aef3ff","workflow_id":"rb-sy-stage-20260928-pilot-dsflash-sol-gpqa-diamond.nvidia/switchyard.auto-routing","benchmark_type":"gpqa_diamond","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-28T14:41:42.078Z","resumed_from_workflow_id":null,"config_key":"86cc6903ee839ec73bb8f560d7be41e1","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["deepseek/deepseek-v4.1-flash","openai/gpt-6-sol"],"temperature":0.5,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.934343,"total_questions":198,"correct_answers":185,"total_cost":1.446501,"generation_time_ms":9674512,"duration_ms":4747724,"input_tokens":51739,"output_tokens":297904,"reasoning_tokens":278894,"success":true},{"id":"09b9c480-6c08-42c4-8821-28817b4c0d99","workflow_id":"rb-sy-stage-20260928-pilot-glm-opus-gpqa-diamond.nvidia/switchyard.auto-routing","benchmark_type":"gpqa_diamond","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-28T14:41:42.077Z","resumed_from_workflow_id":null,"config_key":"3699c8d71d966e59cd0723fc5ca54716","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["z-ai/glm-5.3-flash","anthropic/claude-opus-5.5"],"temperature":0.5,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.914141,"total_questions":198,"correct_answers":181,"total_cost":3.576265,"generation_time_ms":10308503,"duration_ms":6032641,"input_tokens":67868,"output_tokens":300177,"reasoning_tokens":211489,"success":true},{"id":"e5dc9b2b-9a78-42d7-a1e3-2e3f3c99b561","workflow_id":"rb-sy2-swe-atlas-qa-glm-5.3-e1","benchmark_type":"swe_atlas_qa","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.492Z","resumed_from_workflow_id":null,"config_key":"65da924c80aa72e597bab844f5259b5c","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["z-ai/glm-5.3"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.572581,"total_questions":124,"correct_answers":71,"total_cost":79.881578,"generation_time_ms":67685398,"duration_ms":17550417,"input_tokens":390624350,"output_tokens":3621554,"reasoning_tokens":2029200,"success":true},{"id":"755ac704-b820-46f8-9d83-1deee7c0c897","workflow_id":"rb-sy2-deep-swe-gpt-6-luna-e1","benchmark_type":"deep_swe","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.475Z","resumed_from_workflow_id":null,"config_key":"6b6bd77e89e943b3b23a5ea919aae3d5","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6-luna"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.530973,"total_questions":113,"correct_answers":60,"total_cost":7.500924,"generation_time_ms":46323089,"duration_ms":12458727,"input_tokens":393062965,"output_tokens":4442109,"reasoning_tokens":2624293,"success":true},{"id":"4381aa3b-ef8c-4e56-9deb-384244a87cad","workflow_id":"rb-sy2-tau3-banking-glm-5.3-flash-e1","benchmark_type":"tau3_bench_banking","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.474Z","resumed_from_workflow_id":null,"config_key":"7ad57946a6cf682846fa465e7ab177a4","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["z-ai/glm-5.3-flash"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.371134,"total_questions":97,"correct_answers":36,"total_cost":3.500089,"generation_time_ms":23114942,"duration_ms":3085055,"input_tokens":93464243,"output_tokens":805710,"reasoning_tokens":571913,"success":true},{"id":"1de7c2ab-cb68-4402-b3c1-a26c18f40f6f","workflow_id":"rb-sy2-deep-swe-glm-5.3-e1","benchmark_type":"deep_swe","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.474Z","resumed_from_workflow_id":null,"config_key":"37531a6c4e5fcadb8389f451f129d3b6","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["z-ai/glm-5.3"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.646018,"total_questions":113,"correct_answers":73,"total_cost":165.31288,"generation_time_ms":110230755,"duration_ms":24180261,"input_tokens":805035182,"output_tokens":6760634,"reasoning_tokens":4091951,"success":true},{"id":"499c3fdf-a765-48b6-8edd-7ebc1c7da89a","workflow_id":"rb-sy2-gpqa-gpt-6-sol-e1","benchmark_type":"gpqa_diamond","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.470Z","resumed_from_workflow_id":null,"config_key":"43e4a0bb2111a0207fa609a0b7298f52","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6-sol"],"temperature":0.5,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.914141,"total_questions":198,"correct_answers":181,"total_cost":1.680313,"generation_time_ms":2459868,"duration_ms":830422,"input_tokens":51603,"output_tokens":157569,"reasoning_tokens":142895,"success":true},{"id":"f893c2ab-ab28-4abd-9987-5becf1b36e40","workflow_id":"rb-sy2-gpqa-gpt-6-luna-e1","benchmark_type":"gpqa_diamond","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.469Z","resumed_from_workflow_id":null,"config_key":"77f89a9d4b1d3ffc9e3dbc2e6340ae0f","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6-luna"],"temperature":0.5,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.868687,"total_questions":198,"correct_answers":172,"total_cost":0.194184,"generation_time_ms":3480886,"duration_ms":845517,"input_tokens":51603,"output_tokens":377906,"reasoning_tokens":361408,"success":true},{"id":"ca155568-55e9-43fb-af49-ece5d0b7bfcd","workflow_id":"rb-sy2-deep-swe-kimi-k3-e1","benchmark_type":"deep_swe","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.467Z","resumed_from_workflow_id":null,"config_key":"3adcee5e07b29bec643c70c9f65d2477","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["moonshotai/kimi-k3"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.690265,"total_questions":113,"correct_answers":78,"total_cost":209.421451,"generation_time_ms":156955794,"duration_ms":27825503,"input_tokens":531865780,"output_tokens":6374976,"reasoning_tokens":4033413,"success":true},{"id":"c2b5729c-774a-4436-a7ce-3dba6b62f0f2","workflow_id":"rb-sy2-deep-swe-opus-5.5-e1","benchmark_type":"deep_swe","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.467Z","resumed_from_workflow_id":null,"config_key":"4269908affdb9f3e6b24d7a9f81559d2","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["anthropic/claude-opus-5.5"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.734513,"total_questions":113,"correct_answers":83,"total_cost":285.011051,"generation_time_ms":68572475,"duration_ms":11821433,"input_tokens":466458925,"output_tokens":6401253,"reasoning_tokens":2630333,"success":true},{"id":"f8be8bb3-1d57-48b2-ae31-bea7bbbac01f","workflow_id":"rb-sy2-tau3-banking-kimi-k3-e1","benchmark_type":"tau3_bench_banking","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.466Z","resumed_from_workflow_id":null,"config_key":"58e1a20ebc39b2870924bbeedfc4a2a4","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["moonshotai/kimi-k3"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.381443,"total_questions":97,"correct_answers":37,"total_cost":79.256036,"generation_time_ms":23953672,"duration_ms":3661074,"input_tokens":174668796,"output_tokens":1352334,"reasoning_tokens":1055782,"success":true},{"id":"7f942fdb-9cbd-4476-8493-ea594a1ddde2","workflow_id":"rb-sy2-deep-swe-glm-5.3-flash-e1","benchmark_type":"deep_swe","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.466Z","resumed_from_workflow_id":null,"config_key":"c5be45086cc4476cd096912811ecd0a5","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["z-ai/glm-5.3-flash"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.309735,"total_questions":113,"correct_answers":35,"total_cost":38.67259,"generation_time_ms":183324932,"duration_ms":43380141,"input_tokens":1346445400,"output_tokens":7552682,"reasoning_tokens":4354715,"success":true},{"id":"349408f7-d74d-4e82-9c98-87ff66919e43","workflow_id":"rb-sy2-gpqa-glm-5.3-e1","benchmark_type":"gpqa_diamond","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.466Z","resumed_from_workflow_id":null,"config_key":"bc593430cc04c9fcac9b48f571c6b6ff","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["z-ai/glm-5.3"],"temperature":0.5,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.853535,"total_questions":198,"correct_answers":169,"total_cost":3.918581,"generation_time_ms":21608838,"duration_ms":3135642,"input_tokens":52757,"output_tokens":2204128,"reasoning_tokens":2139791,"success":true},{"id":"ab45d8fc-8a1e-4762-b59d-70b43382c944","workflow_id":"rb-sy2-swe-atlas-qa-gpt-6-luna-e1","benchmark_type":"swe_atlas_qa","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.465Z","resumed_from_workflow_id":null,"config_key":"7546a08366221a1286545dfc909dbedd","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6-luna"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.354839,"total_questions":124,"correct_answers":44,"total_cost":3.378771,"generation_time_ms":20565552,"duration_ms":6814409,"input_tokens":123903528,"output_tokens":1960772,"reasoning_tokens":1114253,"success":true},{"id":"74de37d5-a739-4836-b3d7-293d5086ecd8","workflow_id":"rb-sy2-deep-swe-gpt-6-astra-e1","benchmark_type":"deep_swe","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.465Z","resumed_from_workflow_id":null,"config_key":"8b94f5faaf143d2deacbec01a3046dff","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6-astra"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.716814,"total_questions":113,"correct_answers":81,"total_cost":303.860228,"generation_time_ms":52143513,"duration_ms":11324406,"input_tokens":92872741,"output_tokens":2581959,"reasoning_tokens":808141,"success":true},{"id":"e6dab572-4300-4544-b46a-dbef2df7219a","workflow_id":"rb-sy2-gpqa-kimi-k3-e1","benchmark_type":"gpqa_diamond","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.463Z","resumed_from_workflow_id":null,"config_key":"4f2d8ebd0d4c506317388e2708686ab9","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["moonshotai/kimi-k3"],"temperature":0.5,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.949495,"total_questions":198,"correct_answers":188,"total_cost":7.091439,"generation_time_ms":8064670,"duration_ms":1835620,"input_tokens":69209,"output_tokens":650009,"reasoning_tokens":617249,"success":true},{"id":"317076e4-c08f-47b3-92ba-854767ea15f4","workflow_id":"rb-sy2-tau3-banking-glm-5.3-e1","benchmark_type":"tau3_bench_banking","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.463Z","resumed_from_workflow_id":null,"config_key":"f42c85bdbd6833935dd40652771dd34d","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["z-ai/glm-5.3"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.381443,"total_questions":97,"correct_answers":37,"total_cost":18.243153,"generation_time_ms":11102490,"duration_ms":1704176,"input_tokens":73461209,"output_tokens":594469,"reasoning_tokens":374071,"success":true},{"id":"0e74d1e5-4fba-488e-90c0-9b2652b87266","workflow_id":"rb-sy2-swe-atlas-qa-gpt-6-astra-e1","benchmark_type":"swe_atlas_qa","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.463Z","resumed_from_workflow_id":null,"config_key":"386401ae0243becfb10e7473827318b1","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6-astra"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.655462,"total_questions":119,"correct_answers":78,"total_cost":244.95841,"generation_time_ms":34152701,"duration_ms":6629183,"input_tokens":75023099,"output_tokens":1575736,"reasoning_tokens":330588,"success":true},{"id":"d8210edd-438c-4d39-a2cb-74dc156bb386","workflow_id":"rb-sy2-swe-atlas-qa-glm-5.3-flash-e1","benchmark_type":"swe_atlas_qa","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.462Z","resumed_from_workflow_id":null,"config_key":"094bac727a85f033d47ef20ec62e5211","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["z-ai/glm-5.3-flash"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.427419,"total_questions":124,"correct_answers":53,"total_cost":24.573037,"generation_time_ms":112241756,"duration_ms":32969291,"input_tokens":758829338,"output_tokens":3334940,"reasoning_tokens":1147003,"success":true},{"id":"57af040d-f063-4545-b2b4-d48808865d47","workflow_id":"rb-sy2-gpqa-gpt-6-astra-e1","benchmark_type":"gpqa_diamond","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.462Z","resumed_from_workflow_id":null,"config_key":"2ad41862447eb9db9da44cb280062fb6","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6-astra"],"temperature":0.5,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.949495,"total_questions":198,"correct_answers":188,"total_cost":6.700163,"generation_time_ms":3172392,"duration_ms":783251,"input_tokens":51603,"output_tokens":123541,"reasoning_tokens":89783,"success":true},{"id":"8adc4422-6c33-437c-ac67-3f59c70678fc","workflow_id":"rb-sy2-swe-atlas-qa-opus-5.5-e1","benchmark_type":"swe_atlas_qa","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.461Z","resumed_from_workflow_id":null,"config_key":"a8610eb8c38c2a652164d11318f98b22","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["anthropic/claude-opus-5.5"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.556452,"total_questions":124,"correct_answers":69,"total_cost":98.336004,"generation_time_ms":28300527,"duration_ms":10123391,"input_tokens":134632195,"output_tokens":1954062,"reasoning_tokens":479825,"success":true},{"id":"cbc48863-9f93-4a30-b412-5ce65e506af3","workflow_id":"rb-sy2-swe-atlas-qa-gpt-6-sol-e1","benchmark_type":"swe_atlas_qa","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.460Z","resumed_from_workflow_id":null,"config_key":"7564c56b45ff6a63e21f296ca0ca67a3","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6-sol"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.553719,"total_questions":121,"correct_answers":67,"total_cost":42.648324,"generation_time_ms":19102401,"duration_ms":6615331,"input_tokens":81392889,"output_tokens":1072116,"reasoning_tokens":239787,"success":true},{"id":"b12fd7bd-e760-476d-8ac1-19cafb752451","workflow_id":"rb-sy2-tau3-banking-gpt-6-astra-e1","benchmark_type":"tau3_bench_banking","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.460Z","resumed_from_workflow_id":null,"config_key":"4cf1fc0452f66150c088ad95d6c9ca16","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6-astra"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.391753,"total_questions":97,"correct_answers":38,"total_cost":181.777802,"generation_time_ms":12312741,"duration_ms":1791444,"input_tokens":93104908,"output_tokens":335460,"reasoning_tokens":133169,"success":true},{"id":"a4eab3ca-57bd-4808-9fd2-a1be3fdc703c","workflow_id":"rb-sy2-tau3-banking-gpt-6-luna-e1","benchmark_type":"tau3_bench_banking","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.460Z","resumed_from_workflow_id":null,"config_key":"97274dcbc07fcfbd764664c6cdd25c6c","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6-luna"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.247423,"total_questions":97,"correct_answers":24,"total_cost":1.964327,"generation_time_ms":10077332,"duration_ms":1558428,"input_tokens":86079711,"output_tokens":852326,"reasoning_tokens":683991,"success":true},{"id":"205cab7b-cb83-4840-a668-4762f21028bc","workflow_id":"rb-sy2-tau3-banking-opus-5.5-e1","benchmark_type":"tau3_bench_banking","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.460Z","resumed_from_workflow_id":null,"config_key":"443b674d7a7c693e085aab52b11af045","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["anthropic/claude-opus-5.5"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.42268,"total_questions":97,"correct_answers":41,"total_cost":71.171822,"generation_time_ms":12229173,"duration_ms":1756792,"input_tokens":121420226,"output_tokens":590819,"reasoning_tokens":191391,"success":true},{"id":"57ca9207-1524-45f7-b476-3e422bbde67a","workflow_id":"rb-sy2-gpqa-glm-5.3-flash-e1","benchmark_type":"gpqa_diamond","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.457Z","resumed_from_workflow_id":null,"config_key":"4e9e5e4226a4d488d18291c89af0e7bf","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["z-ai/glm-5.3-flash"],"temperature":0.5,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.893401,"total_questions":197,"correct_answers":176,"total_cost":0.588095,"generation_time_ms":57948853,"duration_ms":13269456,"input_tokens":54714,"output_tokens":1906995,"reasoning_tokens":1857813,"success":true},{"id":"fa77744f-616d-4566-a946-c19c9a86a08f","workflow_id":"rb-sy2-swe-atlas-qa-kimi-k3-e1","benchmark_type":"swe_atlas_qa","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.436Z","resumed_from_workflow_id":null,"config_key":"3a04ee35e708ab56870b65a1fc96950b","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["moonshotai/kimi-k3"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.580645,"total_questions":124,"correct_answers":72,"total_cost":107.139966,"generation_time_ms":60598317,"duration_ms":15396268,"input_tokens":239186689,"output_tokens":2934668,"reasoning_tokens":1567034,"success":true},{"id":"4fb0f7de-9f30-4ed2-a977-75aac8754210","workflow_id":"rb-sy2-gpqa-opus-5.5-e1","benchmark_type":"gpqa_diamond","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.436Z","resumed_from_workflow_id":null,"config_key":"f19307fecfab8ab101847cf6f278e8cf","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["anthropic/claude-opus-5.5"],"temperature":0.5,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.893939,"total_questions":198,"correct_answers":177,"total_cost":3.559004,"generation_time_ms":1977303,"duration_ms":711367,"input_tokens":66672,"output_tokens":164011,"reasoning_tokens":66896,"success":true},{"id":"492f9cf9-1019-4d5f-b586-0db79ed58044","workflow_id":"rb-sy2-deep-swe-gpt-6-sol-e1","benchmark_type":"deep_swe","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.431Z","resumed_from_workflow_id":null,"config_key":"c38d0de954cd7d54cd19d45520387d49","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6-sol"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.575221,"total_questions":113,"correct_answers":65,"total_cost":53.744126,"generation_time_ms":29321465,"duration_ms":6182622,"input_tokens":106719256,"output_tokens":1884000,"reasoning_tokens":569472,"success":true},{"id":"96ba2886-4b28-4927-88ae-412bc80e267d","workflow_id":"rb-sy2-tau3-banking-gpt-6-sol-e1","benchmark_type":"tau3_bench_banking","model_permaslug":"nvidia/switchyard","provider_name":"auto-routing","timestamp":"2026-09-23T21:40:42.429Z","resumed_from_workflow_id":null,"config_key":"4ea33e0f90ee1458d47077dfe9ad0c90","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":["openai/gpt-6-sol"],"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.42268,"total_questions":97,"correct_answers":41,"total_cost":52.904438,"generation_time_ms":9137863,"duration_ms":1468524,"input_tokens":137742468,"output_tokens":437920,"reasoning_tokens":226743,"success":true},{"id":"a88b0560-94d9-43aa-a331-cfd22320d324","workflow_id":"rb-7741ebdf-ext5b-fugu-max-deep-swe","benchmark_type":"deep_swe","model_permaslug":"sakana/fugu-max","provider_name":"auto-routing","timestamp":"2026-09-18T17:14:33.283Z","resumed_from_workflow_id":null,"config_key":"966cf9b20f1df4816400a69d5baf59dd","reasoning_effort":"medium","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":2,"recorded_epochs":2,"accuracy":0.339286,"total_questions":224,"correct_answers":76,"total_cost":672.333734,"generation_time_ms":341844679,"duration_ms":62962733,"input_tokens":1811999956,"output_tokens":26015609,"reasoning_tokens":19580403,"success":true},{"id":"da7352d2-7e87-49c3-9df5-70b71c14dcee","workflow_id":"rb-7741ebdf-ext5b-fugu-ultra-v2-swe-atlas-qa","benchmark_type":"swe_atlas_qa","model_permaslug":"sakana/fugu-ultra-v2","provider_name":"auto-routing","timestamp":"2026-09-18T17:14:33.223Z","resumed_from_workflow_id":null,"config_key":"a26832f83753b3092f51e30f4cda26a0","reasoning_effort":"medium","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.694915,"total_questions":118,"correct_answers":82,"total_cost":1051.581245,"generation_time_ms":109569714,"duration_ms":19021856,"input_tokens":568897814,"output_tokens":7444848,"reasoning_tokens":2202523,"success":true},{"id":"e654a444-a0ff-4307-b763-2d1c47cfc36c","workflow_id":"rb-7741ebdf-ext4-fugu-max-mmmu-pro-vision","benchmark_type":"mmmu_pro_vision","model_permaslug":"sakana/fugu-max","provider_name":"auto-routing","timestamp":"2026-09-18T11:55:14.198Z","resumed_from_workflow_id":null,"config_key":"c7c88655aafd2ff3f66e9a009ddf0b1f","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":3,"recorded_epochs":3,"accuracy":0.794605,"total_questions":5190,"correct_answers":4124,"total_cost":100.348075,"generation_time_ms":126036381,"duration_ms":7168366,"input_tokens":8029093,"output_tokens":14302323,"reasoning_tokens":13288841,"success":true},{"id":"30d03ff8-777a-475f-b04d-fb73ae05785b","workflow_id":"rb-7741ebdf-ext4-fugu-max-tau3-banking","benchmark_type":"tau3_bench_banking","model_permaslug":"sakana/fugu-max","provider_name":"auto-routing","timestamp":"2026-09-18T11:55:14.193Z","resumed_from_workflow_id":null,"config_key":"631d44e09067c0e58125bd89a32ee325","reasoning_effort":"medium","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.298969,"total_questions":97,"correct_answers":29,"total_cost":57.555494,"generation_time_ms":19408156,"duration_ms":3587601,"input_tokens":104492046,"output_tokens":1491398,"reasoning_tokens":1224420,"success":true},{"id":"fc4aa9e6-eaaf-4cf2-9cc7-62e21940f167","workflow_id":"rb-7741ebdf-ext4-fugu-ultra-v2-tau3-banking","benchmark_type":"tau3_bench_banking","model_permaslug":"sakana/fugu-ultra-v2","provider_name":"auto-routing","timestamp":"2026-09-18T11:55:14.190Z","resumed_from_workflow_id":null,"config_key":"87ecead7895f766c6093ecc3a56c2f10","reasoning_effort":"medium","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.350515,"total_questions":97,"correct_answers":34,"total_cost":275.673304,"generation_time_ms":23940758,"duration_ms":3608817,"input_tokens":131713843,"output_tokens":1327595,"reasoning_tokens":802896,"success":true},{"id":"672b44c4-0d7c-41b5-92e7-17c2066cbdab","workflow_id":"rb-7741ebdf-ext4-fugu-ultra-v2-mmmu-pro-vision","benchmark_type":"mmmu_pro_vision","model_permaslug":"sakana/fugu-ultra-v2","provider_name":"auto-routing","timestamp":"2026-09-18T11:55:14.184Z","resumed_from_workflow_id":null,"config_key":"a10543c341a126603b5856b3a301ba50","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":3,"recorded_epochs":3,"accuracy":0.80501,"total_questions":5190,"correct_answers":4178,"total_cost":567.894964,"generation_time_ms":218281287,"duration_ms":9977928,"input_tokens":85928291,"output_tokens":5902077,"reasoning_tokens":587726,"success":true},{"id":"a511c523-45a5-4e63-9e2d-6ddb22fc3fd6","workflow_id":"rb-7741ebdf-ext4-fugu-max-swe-atlas-qa","benchmark_type":"swe_atlas_qa","model_permaslug":"sakana/fugu-max","provider_name":"auto-routing","timestamp":"2026-09-18T11:55:13.783Z","resumed_from_workflow_id":null,"config_key":"392fe58ae6feb02dc5fca40e1e66a234","reasoning_effort":"medium","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.524194,"total_questions":124,"correct_answers":65,"total_cost":193.8577,"generation_time_ms":82553717,"duration_ms":19650054,"input_tokens":530337376,"output_tokens":6213915,"reasoning_tokens":4028249,"success":true},{"id":"0fa6f6e3-5dc1-48cc-af42-0d6461f94cda","workflow_id":"rb-7741ebdf-ext-fugu-max-deep-swe","benchmark_type":"deep_swe","model_permaslug":"sakana/fugu-max","provider_name":"auto-routing","timestamp":"2026-09-16T03:06:09.565Z","resumed_from_workflow_id":null,"config_key":"966cf9b20f1df4816400a69d5baf59dd","reasoning_effort":"medium","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.383929,"total_questions":112,"correct_answers":43,"total_cost":322.464822,"generation_time_ms":166433486,"duration_ms":44515180,"input_tokens":855890840,"output_tokens":13167225,"reasoning_tokens":10921808,"success":true},{"id":"82ffd18b-05b5-41f4-b0b4-29edb12978cb","workflow_id":"rb-7741ebdf-ext-fugu-max-swe-atlas-qa","benchmark_type":"swe_atlas_qa","model_permaslug":"sakana/fugu-max","provider_name":"auto-routing","timestamp":"2026-09-16T02:46:05.137Z","resumed_from_workflow_id":null,"config_key":"392fe58ae6feb02dc5fca40e1e66a234","reasoning_effort":"medium","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":2,"recorded_epochs":2,"accuracy":0.528226,"total_questions":248,"correct_answers":131,"total_cost":406.188523,"generation_time_ms":130083478,"duration_ms":22825043,"input_tokens":1045289588,"output_tokens":13086401,"reasoning_tokens":8675369,"success":true},{"id":"cf5eaf24-3fa3-4ca3-9c3c-e1ee7387e810","workflow_id":"rb-7741ebdf-ext-fugu-max-tau3-banking","benchmark_type":"tau3_bench_banking","model_permaslug":"sakana/fugu-max","provider_name":"auto-routing","timestamp":"2026-09-16T02:46:05.133Z","resumed_from_workflow_id":null,"config_key":"631d44e09067c0e58125bd89a32ee325","reasoning_effort":"medium","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":2,"recorded_epochs":2,"accuracy":0.304124,"total_questions":194,"correct_answers":59,"total_cost":84.783657,"generation_time_ms":23498566,"duration_ms":3122623,"input_tokens":139764660,"output_tokens":2370109,"reasoning_tokens":1945169,"success":true},{"id":"a5a585d4-4756-4390-ada5-22a5be25e79e","workflow_id":"rb-7741ebdf-ext-fugu-ultra-v2-tau3-banking","benchmark_type":"tau3_bench_banking","model_permaslug":"sakana/fugu-ultra-v2","provider_name":"auto-routing","timestamp":"2026-09-16T02:46:05.133Z","resumed_from_workflow_id":null,"config_key":"87ecead7895f766c6093ecc3a56c2f10","reasoning_effort":"medium","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":2,"recorded_epochs":2,"accuracy":0.345361,"total_questions":194,"correct_answers":67,"total_cost":628.777126,"generation_time_ms":51402317,"duration_ms":4308575,"input_tokens":303783938,"output_tokens":3071479,"reasoning_tokens":1901712,"success":true},{"id":"e95ba7bb-a708-44d8-918c-ab60aab3692f","workflow_id":"rb-7741ebdf-pilot-auto-search-hle","benchmark_type":"search_hle","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-15T18:51:27.288Z","resumed_from_workflow_id":null,"config_key":"e15d82a9092d75202869b7b3709fc905","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":1,"recorded_epochs":1,"accuracy":0.524503,"total_questions":2061,"correct_answers":1081,"total_cost":168.79506,"generation_time_ms":254008292,"duration_ms":72025759,"input_tokens":263870077,"output_tokens":36394607,"reasoning_tokens":33677094,"success":true},{"id":"d41c0935-de23-45d3-a42a-a82e4e1d3fe6","workflow_id":"rb-7741ebdf-pilot-auto-mmmu-pro-vision","benchmark_type":"mmmu_pro_vision","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-15T18:49:01.492Z","resumed_from_workflow_id":null,"config_key":"cc08532109262d3bf45695d00da9cf72","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":3,"recorded_epochs":3,"accuracy":0.744316,"total_questions":5190,"correct_answers":3863,"total_cost":32.949705,"generation_time_ms":140726769,"duration_ms":10495867,"input_tokens":16233961,"output_tokens":14752498,"reasoning_tokens":13136167,"success":true},{"id":"b31d55ff-2786-464c-a413-83c8104cef39","workflow_id":"rb-7741ebdf-pilot-auto-deep-swe","benchmark_type":"deep_swe","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-15T17:19:36.246Z","resumed_from_workflow_id":null,"config_key":"3fbb66e0a988cd24f2a9651673c8bff1","reasoning_effort":"medium","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":3,"recorded_epochs":3,"accuracy":0.157817,"total_questions":336,"correct_answers":51,"total_cost":15.32906,"generation_time_ms":58372403,"duration_ms":16573748,"input_tokens":364344616,"output_tokens":3333815,"reasoning_tokens":1032928,"success":true},{"id":"e13c5903-b77a-4d53-9482-b47e3a9b6317","workflow_id":"rb-7741ebdf-pilot-auto-swe-atlas-qa","benchmark_type":"swe_atlas_qa","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-15T14:09:23.266Z","resumed_from_workflow_id":null,"config_key":"a0616a6e6c3e742b67c28b8fc714ec7b","reasoning_effort":"medium","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":3,"recorded_epochs":3,"accuracy":0.287634,"total_questions":372,"correct_answers":107,"total_cost":7.701978,"generation_time_ms":26799842,"duration_ms":6248734,"input_tokens":103438885,"output_tokens":1828664,"reasoning_tokens":250113,"success":true},{"id":"310e45ac-fa81-49dd-93d2-6cd0e5e621ff","workflow_id":"rb-7741ebdf-pilot-auto-gpqa","benchmark_type":"gpqa_diamond","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-15T14:00:47.531Z","resumed_from_workflow_id":null,"config_key":"032aa9b706852f04e602639bc76d02c4","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0.5,"sample_limit":null,"epochs":3,"recorded_epochs":3,"accuracy":0.900673,"total_questions":594,"correct_answers":535,"total_cost":7.072428,"generation_time_ms":32471699,"duration_ms":3245131,"input_tokens":192171,"output_tokens":5413208,"reasoning_tokens":5377855,"success":true},{"id":"005a50a6-f4d1-4067-b5e4-9354fc1a362c","workflow_id":"rb-7741ebdf-pilot-auto-tau3-banking","benchmark_type":"tau3_bench_banking","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-15T14:00:47.340Z","resumed_from_workflow_id":null,"config_key":"9e57d33358f22bc1e824bc85f9520be9","reasoning_effort":"medium","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0,"sample_limit":null,"epochs":3,"recorded_epochs":3,"accuracy":0.137457,"total_questions":291,"correct_answers":40,"total_cost":25.221053,"generation_time_ms":26794842,"duration_ms":3227461,"input_tokens":184872969,"output_tokens":2530498,"reasoning_tokens":2024715,"success":true},{"id":"d8fbce64-1a8e-4648-827c-962d3986cdad","workflow_id":"rbgq3-1789433899-pilot-gpqa.sakana/fugu-ultra-v2.auto-routing","benchmark_type":"gpqa_diamond","model_permaslug":"sakana/fugu-ultra-v2","provider_name":"auto-routing","timestamp":"2026-09-15T01:07:48.052Z","resumed_from_workflow_id":null,"config_key":"37a4dbebefa2842508adf5f32c99406b","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0.5,"sample_limit":null,"epochs":3,"recorded_epochs":3,"accuracy":0.93266,"total_questions":594,"correct_answers":554,"total_cost":99.506314,"generation_time_ms":38593253,"duration_ms":1970735,"input_tokens":5035457,"output_tokens":2733633,"reasoning_tokens":324526,"success":true},{"id":"814ad8a9-7b16-45c4-9bb2-e4cfb657b9b8","workflow_id":"rbgq3-1789433899-pilot-gpqa.sakana/fugu-ultra.auto-routing","benchmark_type":"gpqa_diamond","model_permaslug":"sakana/fugu-ultra","provider_name":"auto-routing","timestamp":"2026-09-15T01:07:47.987Z","resumed_from_workflow_id":null,"config_key":"634d38fa293f5999bdf29b2cb3d47c4b","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0.5,"sample_limit":null,"epochs":3,"recorded_epochs":3,"accuracy":0.946128,"total_questions":594,"correct_answers":562,"total_cost":227.450382,"generation_time_ms":65098262,"duration_ms":3830881,"input_tokens":6522402,"output_tokens":6587778,"reasoning_tokens":706903,"success":true},{"id":"6d78f821-57e3-4378-a725-7c5133979925","workflow_id":"rbgq3-1789433899-pilot-gpqa.openrouter/auto.auto-routing","benchmark_type":"gpqa_diamond","model_permaslug":"openrouter/auto","provider_name":"auto-routing","timestamp":"2026-09-15T01:07:47.971Z","resumed_from_workflow_id":null,"config_key":"90fa47f67f9993814d325aefb0cf9183","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0.5,"sample_limit":null,"epochs":3,"recorded_epochs":3,"accuracy":0.907407,"total_questions":594,"correct_answers":539,"total_cost":1.560001,"generation_time_ms":46948387,"duration_ms":4171740,"input_tokens":163266,"output_tokens":4642534,"reasoning_tokens":4601108,"success":true},{"id":"fc1568ed-851a-44a8-87a2-98139b4ee79f","workflow_id":"rbgq3-1789433899-pilot-gpqa.sakana/fugu-max.auto-routing","benchmark_type":"gpqa_diamond","model_permaslug":"sakana/fugu-max","provider_name":"auto-routing","timestamp":"2026-09-15T01:07:47.957Z","resumed_from_workflow_id":null,"config_key":"7b1f4ced9764072b0c37201de4c6ede4","reasoning_effort":"high","max_tokens":null,"cost_tier":null,"route_models":null,"temperature":0.5,"sample_limit":null,"epochs":3,"recorded_epochs":3,"accuracy":0.885522,"total_questions":594,"correct_answers":526,"total_cost":14.959829,"generation_time_ms":44568206,"duration_ms":4589302,"input_tokens":231931,"output_tokens":2431351,"reasoning_tokens":2685693,"success":true}],"truncated":false}}