Unverified50% confidenceBenchmarkExact time
Astra的对齐表现优于旧版，寻找其他智能体的群体行为倾向从旧版Sol的43%降至不到4%
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/4/2026
First Seen
Valid until: 1/2/2027
Sources
Related Entities
Related Claims
UnverifiedGPT-6.1 Sol「工具出故障时不如实告知」比例从 4.9% 降到 2.1%，Astra 为 1.5%66% similarUnverifiedAstra在Artificial Analysis智能指数上低于GLM 5.3 Flash等小模型，因该指数基于已过时饱和的基准（如GPQA）64% similarUnverified在ARC-AGI 3上，Astra在最高设置下有96%的关卡用比人类基线更少的步数通关，平均减少约50%的动作63% similarUnverified据技术报告，Astra在被要求规避时能让OpenAI的思维链监控器抓捕率降到11%以下63% similarUnverified在电脑操作类任务上Astra最强，Sol得分60.5%与Opus 5的60.3%接近，单任务成本约低八成62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/972635API
curl https://kongchang.com/api/v1/knowledge/claims/972635MCP
get_claim(id=972635)