Unverified50% confidenceFactExact time
The safety reports for both Opus 5.5 and Astra show that the models frequently suspected they were being tested during evaluations.
1
Sources
50%
Confidence
Long-term
Relevance
9/29/2026
First Seen
Sources
Related Entities
Related Claims
Unverified社区基准测试可能无意触发安全分类器，导致实际测试的是被路由的Opus 4.8而非目标模型69% similarUnverifiedOpus 5基本一次性通过了物流系统的RPC权限体系测试，但个别地方出现权限校验瑕疵66% similarUnverifiedAnthropic states that in its most comprehensive alignment testing, Opus 5.5 is the best-performing model it has tested to date.65% similarVerifiedBenchmarks tend to evaluate how much a model knows, rather than how much cognitive increment it can bring.64% similarUnverifiedOpus is testing the ability to similarly switch effort per message without breaking the cache, and is currently in the beta stage of the API.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/958941API
curl https://kongchang.com/api/v1/knowledge/claims/958941MCP
get_claim(id=958941)