{"version":"configuration-source-5","evidencePolicy":"documented","snapshotId":"sha256-df858793dcb753b0f77beb223fce04e1de99a2c08f2c7b5f14534efcafa8c7c5","effortDatasetId":"sha256-1f186994047883a45f76e96fbce2f07320baf4f104506414be2bba69a334a02f+sha256-2bc644d4e2fb34bd4fecd1ef0a2a50302a9d18f3b8cd3aa0e6dfc487669f5146+sha256-b6d2e58f71798704ee9f3e6b341e3627e6b0dc1caef890fc1e08b62a226e3f36+sha256-7a7d95c40c608cc28483fb5cef9007da8c102f1e7278882ccbcc02a16078607d+sha256-99e5ccb2eac4abb63cc6d9823b591f2ffac992cdee03ac725f66bd8a5ec086a3+sha256-9fbe9729fecc9cf9232f3292684307ba1a85d69beff3920a2a7ed9e2c77bf7ef+sha256-7e00d90a437b5a33e79607a04b57b10dc14322bbaa63ce8bbe9ce1338eaf26f5","scope":"All collected observations, not a claim that every public result has been collected. Contributing means positive graph weight, not sufficient evidence for a rank.","models":[{"id":"grok-4.5","collected":67,"contributing":47,"admittedWithoutWeight":0,"excluded":20}],"observations":[{"id":"e04b4c1584c40ead300a6494","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/aa-analyst-agent","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"54926ad11c502455b00573fe","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/aa-briefcase","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"3fb681f994810c739344b77f","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/artificial-analysis-long-context-reasoning","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"fa8a12feceedbdedf24b72f3","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/automationbench-aa","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"5f6cab432b7af151e56d72c7","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/critpt","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"927091b6f30b8898774d9b75","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/enterprise-ops-gym-aa","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"74d938a06c551969af7cf629","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/gdp-pdf","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"8c4189a8314621ca519523e6","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/gdpval-aa","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"cce6b0003a75b934b5ce797d","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/gpqa-diamond","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"26e57c4171d8af90ff7c9d8b","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/harvey-lab-aa","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"dfc2d18c6b5c05d46a2fc38b","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/humanitys-last-exam","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"2b62bae9938cc7ca82751c8a","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/mlcr-aa","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"0d148747c6b768836b3329d1","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/mmmu-pro","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"49ddf105c76cb6c65c9788e6","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/omniscience","status":"excluded","reason":"Benchmark family or source protocol has not been reviewed"},{"id":"d472e7944026d02ff693dcbe","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/omniscience","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"8b158e8057762eda00e49b8a","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/scicode","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"91458be076d7e97fd168922a","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/tau3-banking","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"377fe99391f8f823331313ae","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/terminalbench-v2-1","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"0212dd4561acee977e00400d","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/terminalbench-v4-0","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"3db181a4b2c903d1c6f5ca32","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://arcprize.org/media/data/leaderboard/v2.json","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"6b3763ee781610c75ebd4561","modelId":"grok-4.5","configurationId":"grok-4.5--effort-low","sourceUrl":"https://arcprize.org/media/data/leaderboard/v2.json","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"f3dd688412c9fb5675915351","modelId":"grok-4.5","configurationId":"grok-4.5--effort-medium","sourceUrl":"https://arcprize.org/media/data/leaderboard/v2.json","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"29895fdf00c54de0c01daedb","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://deepswe.datacurve.ai/artifacts/v1.1/leaderboard-live.json","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"43896251f67c4560c987eefb","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://www.tbench.ai/?version=2.1","status":"excluded","reason":"Fewer than two distinct model identities remain in this evaluation"},{"id":"0e654cd1466ae894d75cb1e8","modelId":"grok-4.5","configurationId":"grok-4.5--effort-medium","sourceUrl":"https://vulcanbench.com/benchmarks/07-grok-fable-sol.html","status":"excluded","reason":"Evaluator has not been reviewed for independent admission"},{"id":"ffbbc7256f47ee72d321f094","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://vulcanbench.com/benchmarks/07-grok-fable-sol.html","status":"excluded","reason":"Evaluator has not been reviewed for independent admission"},{"id":"55c55b9ae75c2c789bb81880","modelId":"grok-4.5","configurationId":"grok-4.5--effort-low","sourceUrl":"https://vulcanbench.com/benchmarks/07-grok-fable-sol.html","status":"excluded","reason":"Evaluator has not been reviewed for independent admission"},{"id":"2bd0af0abcc178550c64a1ac","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://cognition.com/frontiercode","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"7fe01339271092343301c357","modelId":"grok-4.5","configurationId":"grok-4.5--effort-low","sourceUrl":"https://arcprize.org/results/openai-gpt-6-astra","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"826615a0c654de07c920208e","modelId":"grok-4.5","configurationId":"grok-4.5--effort-low","sourceUrl":"https://cognition.com/frontiercode","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"99e65bc1e02ebf4fee08b548","modelId":"grok-4.5","configurationId":"grok-4.5--effort-medium","sourceUrl":"https://arcprize.org/results/openai-gpt-6-astra","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"b7467edb46036b2935325b6c","modelId":"grok-4.5","configurationId":"grok-4.5--effort-medium","sourceUrl":"https://cognition.com/frontiercode","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"c9b43e92020f700c0d052db9","modelId":"grok-4.5","configurationId":"grok-4.5--effort-low","sourceUrl":"https://cognition.com/frontiercode","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"dd27c86aaec46afa668fb3ed","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://arcprize.org/results/openai-gpt-6-astra","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"de359ceaf316774036c9cd92","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://cognition.com/frontiercode","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"eaf538a9130f43391ba85bbf","modelId":"grok-4.5","configurationId":"grok-4.5--effort-medium","sourceUrl":"https://cognition.com/frontiercode","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"ec5579f64a5a32bfa01b650d","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://snorkel.ai/leaderboard/terminal-bench-4-0/","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"panel-c9c71176cdf68d6328da3f8c","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://swe-rebench.com/","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"epoch-run-6e3LDVuZwoZC4Sg7ushoVP","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://epoch.ai/benchmarks/chess-puzzles","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"epoch-run-7HWW3sjRstKzb9JpcnvLaV","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://epoch.ai/benchmarks/simpleqa-verified","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"epoch-run-7vThbrtbczScg2UcsAYXtF","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://epoch.ai/benchmarks/frontiermath-tier-4-v2","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"epoch-run-9kSgmUd9vh4CBXz3tJhctN","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://epoch.ai/benchmarks/gpqa-diamond","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"epoch-run-djbbEqvBW9qux3pZwcZN2i","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://epoch.ai/benchmarks/frontiermath-tiers-1-3-v2","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"epoch-run-f8Pac6hqib4t7x2s4HCRoT","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://epoch.ai/benchmarks/otis-mock-aime-2024-2025","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"vals-finance-agent-v2:grok/grok-4.5","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://www.vals.ai/benchmarks/fabv2","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"vals-code-migration-v1:grok/grok-4.5","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://www.vals.ai/benchmarks/code-migration","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"vals-legal-research-v1:grok/grok-4.5","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://www.vals.ai/benchmarks/legal_research","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"vals-emb-v1:grok/grok-4.5","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://www.vals.ai/benchmarks/emb","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"vals-mmlu-pro-v1:grok/grok-4.5","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://www.vals.ai/benchmarks/mmlu_pro","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"vals-terminal-bench-2-1-v2.1:grok/grok-4.5","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://www.vals.ai/benchmarks/terminal-bench-2-1","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"vals-vibe-code-bench-v1.1:grok/grok-4.5","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://www.vals.ai/benchmarks/vibe-code","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"vals-legal-agent-benchmark-v1:grok/grok-4.5","modelId":"grok-4.5","configurationId":"grok-4.5--effort-high","sourceUrl":"https://www.vals.ai/benchmarks/hlab","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"launch-xai-1319","modelId":"grok-4.5","configurationId":null,"sourceUrl":"https://x.ai/news/grok-4-6","status":"excluded","reason":"Supporting evidence outside the reviewed capability core"},{"id":"launch-xai-1323","modelId":"grok-4.5","configurationId":null,"sourceUrl":"https://x.ai/news/grok-4-6","status":"excluded","reason":"No reviewed effort for this catalog observation; any separately collected effort measurement is counted separately"},{"id":"launch-xai-1327","modelId":"grok-4.5","configurationId":null,"sourceUrl":"https://x.ai/news/grok-4-6","status":"excluded","reason":"Supporting evidence outside the reviewed capability core"},{"id":"launch-xai-1331","modelId":"grok-4.5","configurationId":null,"sourceUrl":"https://x.ai/news/grok-4-6","status":"excluded","reason":"No reviewed effort for this catalog observation; any separately collected effort measurement is counted separately"},{"id":"launch-xai-1335","modelId":"grok-4.5","configurationId":null,"sourceUrl":"https://x.ai/news/grok-4-6","status":"excluded","reason":"Supporting evidence outside the reviewed capability core"},{"id":"launch-xai-1339","modelId":"grok-4.5","configurationId":null,"sourceUrl":"https://x.ai/news/grok-4-6","status":"excluded","reason":"No reviewed effort for this catalog observation; any separately collected effort measurement is counted separately"},{"id":"launch-xai-1343","modelId":"grok-4.5","configurationId":null,"sourceUrl":"https://x.ai/news/grok-4-6","status":"excluded","reason":"No reviewed effort for this catalog observation; any separately collected effort measurement is counted separately"},{"id":"launch-xai-1347","modelId":"grok-4.5","configurationId":null,"sourceUrl":"https://x.ai/news/grok-4-6","status":"excluded","reason":"Supporting evidence outside the reviewed capability core"},{"id":"launch-xai-1350","modelId":"grok-4.5","configurationId":null,"sourceUrl":"https://x.ai/news/grok-4-6","status":"excluded","reason":"Supporting evidence outside the reviewed capability core"},{"id":"launch-xai-1354","modelId":"grok-4.5","configurationId":null,"sourceUrl":"https://x.ai/news/grok-4-6","status":"excluded","reason":"Supporting evidence outside the reviewed capability core"},{"id":"s-deepswe-grok-4.5-eb88d1c756c0","modelId":"grok-4.5","configurationId":null,"sourceUrl":"https://deepswe.datacurve.ai/","status":"excluded","reason":"No reviewed effort for this catalog observation; any separately collected effort measurement is counted separately"},{"id":"s-lmarena-grok-4.5-465fe8c26c46","modelId":"grok-4.5","configurationId":null,"sourceUrl":"https://lmarena.ai/leaderboard","status":"excluded","reason":"No reviewed effort for this catalog observation; any separately collected effort measurement is counted separately"},{"id":"s-vulcanbench-grok-4.5-07-grok-fable-sol-high-2f58434bbe3d","modelId":"grok-4.5","configurationId":null,"sourceUrl":"https://vulcanbench.com/benchmarks/07-grok-fable-sol.html","status":"excluded","reason":"Supporting evidence outside the reviewed capability core"},{"id":"s-vulcanbench-grok-4.5-07-grok-fable-sol-low-2f58434bbe3d","modelId":"grok-4.5","configurationId":null,"sourceUrl":"https://vulcanbench.com/benchmarks/07-grok-fable-sol.html","status":"excluded","reason":"Supporting evidence outside the reviewed capability core"},{"id":"s-vulcanbench-grok-4.5-07-grok-fable-sol-medium-2f58434bbe3d","modelId":"grok-4.5","configurationId":null,"sourceUrl":"https://vulcanbench.com/benchmarks/07-grok-fable-sol.html","status":"excluded","reason":"Supporting evidence outside the reviewed capability core"}]}