{"benchmark_id":"arc-agi-v2","name":"ARC-AGI v2","parent_benchmark":null,"categories":["reasoning","spatial_reasoning","vision"],"modality":"multimodal","multilingual":false,"max_score":1.0,"language":"en","description":"ARC-AGI-2 is an upgraded benchmark for measuring abstract reasoning and problem-solving abilities in AI systems through visual grid transformation tasks. It evaluates fluid intelligence via input-output grid pairs (1x1 to 30x30) using colored cells (0-9), requiring models to identify underlying transformation rules from demonstration examples and apply them to test cases. Designed to be easy for humans but challenging for AI, focusing on core cognitive abilities like spatial reasoning, pattern recognition, and compositional generalization.","paper_link":"https://arxiv.org/abs/2505.11831","implementation_link":null,"verified":false,"created_at":"2026-05-07T16:53:21.712887+00:00","updated_at":"2026-07-05T18:27:43.160253+00:00","statistics":{"total_models":16,"average_score":0.45149999999999996,"min_score":0.049,"max_score":0.85,"score_stddev":0.26710721942071625,"verified_count":0,"self_reported_count":13},"child_benchmarks":[],"linked_dataset":null,"models":[{"rank":1,"model_id":"gpt-5.5","model_name":"GPT-5.5","organization_id":"openai","organization_name":"OpenAI","organization_country":"US","score":0.85,"normalized_score":0.85,"verified":false,"self_reported":true,"self_reported_source":"https://openai.com/index/introducing-gpt-5-5/","analysis_method":"ARC-AGI-2 (Verified). Reasoning effort xhigh.","verification_date":null,"provider_id":"openai","input_cost_per_million":5.0,"output_cost_per_million":30.0,"context_window":1050000,"announcement_date":"2026-04-23","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":"OpenAI","best_throughput":null,"throughput_provider":"OpenAI","context_provider":"OpenAI"},{"rank":2,"model_id":"gemini-3.1-pro-preview","model_name":"Gemini 3.1 Pro","organization_id":"google","organization_name":"Google","organization_country":"US","score":0.771,"normalized_score":0.771,"verified":false,"self_reported":true,"self_reported_source":"https://deepmind.google/models/evals-methodology/gemini-3-1-pro","analysis_method":"ARC Prize Verified","verification_date":null,"provider_id":"google","input_cost_per_million":2.5,"output_cost_per_million":15.0,"context_window":1048576,"announcement_date":"2026-02-19","param_count":null,"is_open_source":false,"is_new":false,"best_latency":0.6,"latency_provider":"Google","best_throughput":90.0,"throughput_provider":"Google","context_provider":"Google"},{"rank":3,"model_id":"gpt-5.4","model_name":"GPT-5.4","organization_id":"openai","organization_name":"OpenAI","organization_country":"US","score":0.733,"normalized_score":0.733,"verified":false,"self_reported":true,"self_reported_source":"https://openai.com/index/introducing-gpt-5-4/","analysis_method":"ARC-AGI-2 (Verified). Reasoning effort xhigh.","verification_date":null,"provider_id":"openai","input_cost_per_million":2.5,"output_cost_per_million":15.0,"context_window":1000000,"announcement_date":"2026-03-05","param_count":null,"is_open_source":false,"is_new":false,"best_latency":3.0,"latency_provider":"OpenAI","best_throughput":50.0,"throughput_provider":"OpenAI","context_provider":"OpenAI"},{"rank":4,"model_id":"gemini-3.5-flash","model_name":"Gemini 3.5 Flash","organization_id":"google","organization_name":"Google","organization_country":"US","score":0.721,"normalized_score":0.721,"verified":false,"self_reported":true,"self_reported_source":"https://deepmind.google/models/evals-methodology/gemini-3-5-flash/","analysis_method":null,"verification_date":null,"provider_id":"google","input_cost_per_million":1.5,"output_cost_per_million":9.0,"context_window":1048576,"announcement_date":"2026-05-19","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":"Google","best_throughput":null,"throughput_provider":"Google","context_provider":"Google"},{"rank":5,"model_id":"claude-opus-4-6","model_name":"Claude Opus 4.6","organization_id":"anthropic","organization_name":"Anthropic","organization_country":"US","score":0.688,"normalized_score":0.688,"verified":false,"self_reported":true,"self_reported_source":"https://www.anthropic.com/news/claude-opus-4-6","analysis_method":"Max effort, 120k thinking budget.","verification_date":null,"provider_id":"anthropic","input_cost_per_million":5.0,"output_cost_per_million":25.0,"context_window":1000000,"announcement_date":"2026-02-05","param_count":null,"is_open_source":false,"is_new":false,"best_latency":0.5,"latency_provider":"Anthropic","best_throughput":42.0,"throughput_provider":"Anthropic","context_provider":"Anthropic"},{"rank":6,"model_id":"claude-sonnet-4-6","model_name":"Claude Sonnet 4.6","organization_id":"anthropic","organization_name":"Anthropic","organization_country":"US","score":0.583,"normalized_score":0.583,"verified":false,"self_reported":true,"self_reported_source":"https://www.anthropic.com/news/claude-sonnet-4-6","analysis_method":"Novel problem-solving (ARC-AGI-2).","verification_date":null,"provider_id":"anthropic","input_cost_per_million":3.0,"output_cost_per_million":15.0,"context_window":200000,"announcement_date":"2026-02-17","param_count":null,"is_open_source":false,"is_new":false,"best_latency":0.4,"latency_provider":"Anthropic","best_throughput":42.0,"throughput_provider":"Anthropic","context_provider":"Anthropic"},{"rank":7,"model_id":"gpt-5.2-pro-2025-12-11","model_name":"GPT-5.2 Pro","organization_id":"openai","organization_name":"OpenAI","organization_country":"US","score":0.542,"normalized_score":0.542,"verified":false,"self_reported":true,"self_reported_source":"https://openai.com/index/introducing-gpt-5-2/","analysis_method":"GPT-5.2 Pro - ARC-AGI-2 (Verified, high compute).","verification_date":null,"provider_id":null,"input_cost_per_million":null,"output_cost_per_million":null,"context_window":null,"announcement_date":"2025-12-11","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":null,"best_throughput":null,"throughput_provider":null,"context_provider":null},{"rank":8,"model_id":"gpt-5.2-2025-12-11","model_name":"GPT-5.2","organization_id":"openai","organization_name":"OpenAI","organization_country":"US","score":0.529,"normalized_score":0.529,"verified":false,"self_reported":true,"self_reported_source":"https://openai.com/index/introducing-gpt-5-2/","analysis_method":"GPT-5.2 Thinking - ARC-AGI-2 (Verified).","verification_date":null,"provider_id":"openai","input_cost_per_million":1.75,"output_cost_per_million":14.0,"context_window":400000,"announcement_date":"2025-12-11","param_count":null,"is_open_source":false,"is_new":false,"best_latency":2.0,"latency_provider":"OpenAI","best_throughput":100.0,"throughput_provider":"OpenAI","context_provider":"OpenAI"},{"rank":9,"model_id":"muse-spark","model_name":"Muse Spark","organization_id":"meta","organization_name":"Meta","organization_country":"US","score":0.425,"normalized_score":0.425,"verified":false,"self_reported":true,"self_reported_source":"https://ai.meta.com/static-resource/muse-spark-eval-methodology","analysis_method":"Thinking mode, pass@2, public set","verification_date":null,"provider_id":null,"input_cost_per_million":null,"output_cost_per_million":null,"context_window":null,"announcement_date":"2026-04-08","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":null,"best_throughput":null,"throughput_provider":null,"context_provider":null},{"rank":10,"model_id":"claude-opus-4-5-20251101","model_name":"Claude Opus 4.5","organization_id":"anthropic","organization_name":"Anthropic","organization_country":"US","score":0.376,"normalized_score":0.376,"verified":false,"self_reported":true,"self_reported_source":"https://www.anthropic.com/news/claude-opus-4-5","analysis_method":"64K thinking budget, interleaved scratchpads, 200K context window, default effort (high).","verification_date":null,"provider_id":null,"input_cost_per_million":null,"output_cost_per_million":null,"context_window":null,"announcement_date":"2025-11-24","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":null,"best_throughput":null,"throughput_provider":null,"context_provider":null},{"rank":11,"model_id":"gemini-3-flash-preview","model_name":"Gemini 3 Flash","organization_id":"google","organization_name":"Google","organization_country":"US","score":0.336,"normalized_score":0.336,"verified":false,"self_reported":true,"self_reported_source":"https://blog.google/products/gemini/gemini-3-flash/","analysis_method":"ARC Prize Verified","verification_date":null,"provider_id":"google","input_cost_per_million":0.5,"output_cost_per_million":3.0,"context_window":1000000,"announcement_date":"2025-12-17","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":"Google","best_throughput":null,"throughput_provider":"Google","context_provider":"Google"},{"rank":12,"model_id":"gemini-3-pro-preview","model_name":"Gemini 3 Pro","organization_id":"google","organization_name":"Google","organization_country":"US","score":0.311,"normalized_score":0.311,"verified":false,"self_reported":true,"self_reported_source":"https://blog.google/products/gemini/gemini-3","analysis_method":"ARC Prize Verified","verification_date":null,"provider_id":null,"input_cost_per_million":null,"output_cost_per_million":null,"context_window":null,"announcement_date":"2025-11-18","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":null,"best_throughput":null,"throughput_provider":null,"context_provider":null},{"rank":13,"model_id":"grok-4","model_name":"Grok-4","organization_id":"xai","organization_name":"xAI","organization_country":"US","score":0.159,"normalized_score":0.159,"verified":false,"self_reported":true,"self_reported_source":"https://x.com/xai/status/1943158495588815072","analysis_method":"accuracy","verification_date":null,"provider_id":null,"input_cost_per_million":null,"output_cost_per_million":null,"context_window":null,"announcement_date":"2025-07-09","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":null,"best_throughput":null,"throughput_provider":null,"context_provider":null},{"rank":14,"model_id":"claude-opus-4-20250514","model_name":"Claude Opus 4","organization_id":"anthropic","organization_name":"Anthropic","organization_country":"US","score":0.086,"normalized_score":0.086,"verified":false,"self_reported":false,"self_reported_source":"https://x.com/xai/status/1943158495588815072","analysis_method":"accuracy","verification_date":null,"provider_id":null,"input_cost_per_million":null,"output_cost_per_million":null,"context_window":null,"announcement_date":"2025-05-22","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":null,"best_throughput":null,"throughput_provider":null,"context_provider":null},{"rank":15,"model_id":"o3-2025-04-16","model_name":"o3","organization_id":"openai","organization_name":"OpenAI","organization_country":"US","score":0.065,"normalized_score":0.065,"verified":false,"self_reported":false,"self_reported_source":"https://x.com/xai/status/1943158495588815072","analysis_method":"accuracy","verification_date":null,"provider_id":null,"input_cost_per_million":null,"output_cost_per_million":null,"context_window":null,"announcement_date":"2025-04-16","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":null,"best_throughput":null,"throughput_provider":null,"context_provider":null},{"rank":16,"model_id":"gemini-2.5-pro","model_name":"Gemini 2.5 Pro","organization_id":"google","organization_name":"Google","organization_country":"US","score":0.049,"normalized_score":0.049,"verified":false,"self_reported":false,"self_reported_source":"https://x.com/xai/status/1943158495588815072","analysis_method":"accuracy","verification_date":null,"provider_id":"google","input_cost_per_million":1.25,"output_cost_per_million":10.0,"context_window":1048576,"announcement_date":"2025-05-20","param_count":null,"is_open_source":false,"is_new":false,"best_latency":0.7,"latency_provider":"Google","best_throughput":85.0,"throughput_provider":"Google","context_provider":"Google"}]}