Unverified50% confidenceBenchmarkExact time
在ARC-AGI-3新颖问题解决基准上模型得分30.2%，而上一代仅1.5%
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
8/23/2026
First Seen
Valid until: 9/6/2026
Sources
Related Entities
Related Claims
Unverified在ARC-AGI-3基准上，GPT-5.6 Sol得分8%，其他模型普遍不足2%，Anthropic因成本过高未运行Fable76% similarUnverified人类在ARC-AGI测试上的平均得分约为85%，而顶尖AI模型长期徘徊在30%-60%区间66% similarUnverifiedDevin 2.0在SWE-bench基准测试中解决了13.86%的实际编程问题，而此前的AI模型仅为1.96%65% similarUnverifiedARC-AGI-2由AI安全研究员François Chollet设计，早期GPT-4在该测试上得分不足10%，人类平均得分约85%64% similarUnverified风险降低类AI用例虽然只占3.4%，但其中有25%报告了变革性ROI64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/791486API
curl https://kongchang.com/api/v1/knowledge/claims/791486MCP
get_claim(id=791486)