Unverified50% confidenceBenchmarkExact time

Claude Sonnet 5 made wild guesses on 25% of cold cases during the development phase, and as many as 30% on Kaggle.

1

Sources

50%

Confidence

Short-term (~14 days)

Relevance

10/11/2026

First Seen

Valid until: 10/25/2026

Sources

Can LLMs Be Data Detectives? A Benchmark Where Honesty Matters More Than Price

Related Entities

Claude Sonnet 5model

Related Claims

UnverifiedIn SWE-bench testing, the early GPT-4 resolution rate was below 2%, while Claude 3.5 Sonnet broke through 49%.59% similar VerifiedClaude Sonnet 5.5 achieved 70.6% on the Terminal benchmark, compared to only 10.3% for the previous generation.59% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/999559

API

curl https://kongchang.com/api/v1/knowledge/claims/999559

MCP

get_claim(id=999559)