{"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":"o4-mini","collected":23,"contributing":19,"admittedWithoutWeight":0,"excluded":4}],"observations":[{"id":"db9246e68971e76beeba26d2","modelId":"o4-mini","configurationId":"o4-mini--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/aime-2025","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"dcb3ed7e008063d590801660","modelId":"o4-mini","configurationId":"o4-mini--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/artificial-analysis-long-context-reasoning","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"efe3ee5e510a1f9d9309239d","modelId":"o4-mini","configurationId":"o4-mini--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/critpt","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"0129f1e1a13aeb775a53a55c","modelId":"o4-mini","configurationId":"o4-mini--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/gpqa-diamond","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"77b640461cfce8756b234629","modelId":"o4-mini","configurationId":"o4-mini--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/humanitys-last-exam","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"48b638db61a973babc2ae6b7","modelId":"o4-mini","configurationId":"o4-mini--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/ifbench","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"2012c98b7c0b198f7fca6c55","modelId":"o4-mini","configurationId":"o4-mini--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/livecodebench","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"1c31fb5e1455b4c4ab80556f","modelId":"o4-mini","configurationId":"o4-mini--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/math-500","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"2116de3638350f5f74312fb0","modelId":"o4-mini","configurationId":"o4-mini--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/mmlu-pro","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"ed9b2508389eebf3c69deacf","modelId":"o4-mini","configurationId":"o4-mini--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/mmmu-pro","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"ed11d2a6911024075e30eda3","modelId":"o4-mini","configurationId":"o4-mini--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/omniscience","status":"excluded","reason":"Benchmark family or source protocol has not been reviewed"},{"id":"5d14e1f9efd0488dbcec0d2f","modelId":"o4-mini","configurationId":"o4-mini--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/omniscience","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"957e1e7000b2a0aa69fbb665","modelId":"o4-mini","configurationId":"o4-mini--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/tau2-bench","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"7e42ac7bd5a7610c724d7104","modelId":"o4-mini","configurationId":"o4-mini--effort-high","sourceUrl":"https://artificialanalysis.ai/evaluations/terminalbench-hard","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"e74df9fbfd59b581eb76ffa2","modelId":"o4-mini","configurationId":"o4-mini--effort-high","sourceUrl":"https://arcprize.org/media/data/leaderboard/v2.json","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"a03ef69f5f0bf9f19a64b6c1","modelId":"o4-mini","configurationId":"o4-mini--effort-low","sourceUrl":"https://arcprize.org/media/data/leaderboard/v2.json","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"904517858e9c5b0473e556b3","modelId":"o4-mini","configurationId":"o4-mini--effort-medium","sourceUrl":"https://arcprize.org/media/data/leaderboard/v2.json","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"ee9d1ca8d2407e02d504ac6b","modelId":"o4-mini","configurationId":"o4-mini--effort-high","sourceUrl":"https://livecodebench.github.io/performances_generation.json","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"c16742f35a5ca0f1da2cd5d9","modelId":"o4-mini","configurationId":"o4-mini--effort-low","sourceUrl":"https://livecodebench.github.io/performances_generation.json","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"dbd3c0c14e4d3b745599aed7","modelId":"o4-mini","configurationId":"o4-mini--effort-medium","sourceUrl":"https://livecodebench.github.io/performances_generation.json","status":"contributing","reason":"Positive comparison weight in the overall connected graph"},{"id":"s-arc2-o4-mini-efebf61be277","modelId":"o4-mini","configurationId":null,"sourceUrl":"https://arcprize.org/leaderboard","status":"excluded","reason":"No reviewed effort for this catalog observation; any separately collected effort measurement is counted separately"},{"id":"s-lcb-o4-mini-ad3f286332a4","modelId":"o4-mini","configurationId":null,"sourceUrl":"https://livecodebench.github.io/leaderboard.html","status":"excluded","reason":"No reviewed effort for this catalog observation; any separately collected effort measurement is counted separately"},{"id":"s-swev-o4-mini-077d8c0cfcca","modelId":"o4-mini","configurationId":null,"sourceUrl":"https://www.swebench.com/","status":"excluded","reason":"No reviewed effort for this catalog observation; any separately collected effort measurement is counted separately"}]}