[{"family":"other-ai","records":23,"currentBests":4,"uncontestedBests":0,"competitions":[{"id":"lean-eval","name":"LeanEval","records":18,"bests":0},{"id":"swe-bench-pro-leaderboard","name":"SWE-bench Pro Leaderboard","records":1,"bests":0},{"id":"mle-bench","name":"MLE-bench","records":4,"bests":4}],"models":{"Tau (caj.al)":1,"Stealth Model":1,"Aleph Prover(logicalintelligence.com)":2,"Midas":1,"github.com/mlgraham (various models)":1,"MerLean-Prover":1,"EVO (deepthought.com.au)":1,"trureturing":1,"Forge":1,"Ivan Bogatyy":1,"EVO":1,"Forall (Astrio)":1,"Various models ( with human in the loop )":1,"Leanstral 1.4":1,"Community multi-model project":1,"Autoform-Bot":1,"Trellis":1,"Claude-Opus-4.5, Claude-Sonnet-4.5, GPT-5.2-Codex, Gemini-3-Pro-Preview":4},"tools":{"AgentX":1,"Disarray":4},"contributors":["lukerj00","rishistyping","mayorov-m-a","sarjinius","mlgraham","doxtor6","test1-deepthought","AlyciaBHZ","xiaoyangliu-sjtu","antpavzhi","bogatyy","machinelearning2014","nolanlwin","yawara","albertqjiang","mrdouglasny","niketp03","wpegden","AgentX-Spring0123","Disarray"],"highConfidence":22},{"family":"openai","records":78,"currentBests":3,"uncontestedBests":1,"competitions":[{"id":"sutro-problems","name":"Sutro Problems","records":13,"bests":2},{"id":"lean-eval","name":"LeanEval","records":19,"bests":0},{"id":"programbench","name":"ProgramBench","records":6,"bests":1},{"id":"mle-bench","name":"MLE-bench","records":40,"bests":0}],"models":{"Codex":5,"GPT 5.6 sol":4,"GPT-5.6":3,"max gp (GPT-6)":1,"savarin-6-hour-gpt-5.6-speedrun":1,"GPT-5.5":2,"deepseek-v4 and gpt-5.x":1,"mostly ChatGPT 5.6 Sol":1,"GPT-5.6 Sol":1,"OpenAI Codex (GPT-5.6 Sol)":1,"HITL":1,"ChatGPT 5.6 Sol":1,"GPT-5.6 Sol)":1,"GPT-5.6 (human-in-the-loop)":1,"Codex (with human in the loop)":1,"GPT-5.5 (xhigh reasoning)":1,"GPT-5.5 (high reasoning)":1,"GPT-5.4":1,"GPT-5.4 mini":1,"GPT-5 mini":1,"gpt-5-codex":4,"gpt-5":8,"o3":4,"o3 + GPT-4.1":4,"o1-preview":8,"gpt-4o-2024-08-06":12},"tools":{"Codex":8,"Humanifa":4,"Codex 5.6 Sol":1,"BRILLiant":1,"DeepScientist":1,"OpenAI Codex (GPT-6":1,"mini-SWE-agent":6,"Thesis":4,"Operand":4,"R&D-Agent":12,"AIRA-dojo":4,"AIDE":8,"OpenHands":4,"MLAB":4},"contributors":["512-unit MLP","PCA-QDA","Reversible MLP","1NN","Ordered ConvNets","cosminscn","sjbaebae","ZhengyangZhang06","Morgan-Griffiths","ma4ypic4y","hongzhoulin89","savarin","juihuichung","antoinegg1","A-M-Berns","eohjelle","ivwumupy","mccorvie","n-yamaguchi-0729","Cond-t","kimihiro64","LexBrill","LikC1606","lixiang90","matthewjasper","plby","GPT-5.5 (xhigh reasoning)","GPT-5.5 (high reasoning)","GPT-5.5","GPT-5.4","GPT-5.4 mini","GPT-5 mini","Thesis","Operand ensemble","R&D-Agent","AIRA-dojo","AIDE","OpenHands","MLAB"],"highConfidence":70},{"family":"anthropic","records":30,"currentBests":1,"uncontestedBests":2,"competitions":[{"id":"lora-speedrun","name":"LoRA Speedrun","records":3,"bests":0},{"id":"sutro-problems","name":"Sutro Problems","records":5,"bests":1},{"id":"lean-eval","name":"LeanEval","records":11,"bests":0},{"id":"programbench","name":"ProgramBench","records":3,"bests":0},{"id":"mle-bench","name":"MLE-bench","records":8,"bests":0}],"models":{"Claude Fable 5":3,"Claude Fable":1,"Claude Opus 5.5":2,"Claude Fable 5.1":1,"Claude Sonnet 4.6":2,"Claude":1,"Antigravity (Multi-Model Ensemble: Gemini 3.1 Pro, Gemini 3 Flash, Claude 4.6 Sonnet/Opus)":1,"Claude Opus 4.7 (1M context)":2,"GPT-5.6 and Fable 5 (human-in-the-loop)":1,"Opus 4.8) + GPT-5.6-Sol":1,"Fable 5 and GPT-5.6":1,"Claude Fable 5 (Claude Code)":1,"Claude Fable 5 (orchestrating Opus/Sonnet 4.8)":1,"Fable 5 (multi-agent)":1,"OpenAI Codex (GPT-5.6 Sol), Claude Code (Opus 5 and Opus 4.8)":1,"Claude Opus 4.6":1,"Claude Haiku 4.5":1,"Claude-Opus-4.6":4,"claude-3-5-sonnet-20240620":4},"tools":{"UNICO/NOUS: Claude (Fable 5":1,"Claude Opus 4.8":1,"mini-SWE-agent":3,"AIBuildAI":4,"AIDE":4},"contributors":["Saivineeth147","stared","yaroslavvb","QDA","SethTS","cosminscn","daouid","rkirov","jzuiddam","KitaKen1","Solarys431","will1491","adamtopaz","homeowmorphism","ianklatzco","ldct","rigwild","Claude Opus 4.6","Claude Sonnet 4.6","Claude Haiku 4.5","AIBuildAI","AIDE"],"highConfidence":26},{"family":"axiom","records":1,"currentBests":1,"uncontestedBests":0,"competitions":[{"id":"lean-eval","name":"LeanEval","records":1,"bests":1}],"models":{"Axiom Prover (Axiom Math)":1},"tools":{},"contributors":["Vilin97"],"highConfidence":1},{"family":"google","records":49,"currentBests":0,"uncontestedBests":0,"competitions":[{"id":"programbench","name":"ProgramBench","records":1,"bests":0},{"id":"mle-bench","name":"MLE-bench","records":48,"bests":0}],"models":{"Gemini 3 Flash":1,"Gemini-3-Pro-Preview":28,"Gemini-2.5-Pro":16,"Gemini-3-Flash-Preview":4},"tools":{"mini-SWE-agent":1,"Famou-Agent 2.0":8,"MARS+":4,"MLEvolve":4,"PiEvolve":8,"MARS":4,"Leeroo":4,"MLE-STAR-Pro-1.5":4,"Famou-Agent":4,"MLE-STAR-Pro-1.0":4,"LoongFlow":4},"contributors":["Gemini 3 Flash","Famou-Agent 2.0","CAIR MARS+","MLEvolve","PiEvolve","CAIR MARS","Leeroo","CAIR MLE-STAR-Pro-1.5","Famou-Agent","CAIR MLE-STAR-Pro-1.0","LoongFlow"],"highConfidence":49},{"family":"deepseek","records":13,"currentBests":0,"uncontestedBests":0,"competitions":[{"id":"lean-eval","name":"LeanEval","records":1,"bests":0},{"id":"mle-bench","name":"MLE-bench","records":12,"bests":0}],"models":{"DeepSeek V4":1,"Deepseek-V3.2-Speciale":4,"deepseek-r1":8},"tools":{"NEAR AI":1,"ML-Master 2.0":4,"InternAgent":4,"ML-Master":4},"contributors":["skidanovalex","ML-Master 2.0","InternAgent","ML-Master"],"highConfidence":13},{"family":"harmonic","records":7,"currentBests":0,"uncontestedBests":0,"competitions":[{"id":"lean-eval","name":"LeanEval","records":7,"bests":0}],"models":{"Aristotle (Harmonic)":7},"tools":{},"contributors":["LorenzoLuccioli","parabamoghv","sqrt-of-2","JohnEdwardJennings","kim-em","adrianmartir","Parcly-Taxel"],"highConfidence":7},{"family":"meta","records":4,"currentBests":0,"uncontestedBests":0,"competitions":[{"id":"mle-bench","name":"MLE-bench","records":4,"bests":0}],"models":{"llama-3.1-405b-instruct":4},"tools":{"AIDE":4},"contributors":["AIDE"],"highConfidence":4},{"family":"bytedance","records":2,"currentBests":0,"uncontestedBests":0,"competitions":[{"id":"lean-eval","name":"LeanEval","records":2,"bests":0}],"models":{"Seed Prover (ByteDance)":2},"tools":{},"contributors":["GanjinZero","hanwenzhu"],"highConfidence":2},{"family":"xai","records":1,"currentBests":0,"uncontestedBests":0,"competitions":[{"id":"lean-eval","name":"LeanEval","records":1,"bests":0}],"models":{"Grok 4.5":1},"tools":{},"contributors":["rwst"],"highConfidence":1}]