Unverified50% confidenceBenchmarkExact time
In the DeepSWE test, 99.6% of all input tokens were cache hits.
1
Sources
50%
Confidence
Long-term
Relevance
9/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn the DeepSWE benchmark, the ratio of input tokens to output tokens during agent execution reached 174 to 1.72% similarUnverifiedThe DeepSuite benchmark has a false positive rate of 44.9% and a false negative rate of 1.2%.68% similarUnverifiedIn the blogger's sample, approximately 94% of Astra's input and approximately 79% of Sonnet's input came from cache.61% similarUnverifiedWork such as LLMLingua from Microsoft Research has demonstrated that removing redundant tokens can compress input tokens by 40%–60% while maintaining task accuracy above 95%.61% similarUnverifiedDeepSeek's hallucination-related score is only about 9%, while GLM 5.3 Flash scores as high as 72% (higher is better).60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/926822API
curl https://kongchang.com/api/v1/knowledge/claims/926822MCP
get_claim(id=926822)