Unverified50% confidenceFactExact time
Nathan Benaich's experiment shows that in human-judged research path divergence experiments, Anthropic's old models could win about 10% of the time, while Mythos-level models had a win rate of about 60%.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/16/2026
First Seen
Valid until: 1/14/2027
Sources
Related Entities
Related Claims
UnverifiedThe paper reports that hybrid reasoning agents achieve accuracy about 10 percentage points higher than conventional models on multiple test sets.67% similarUnverifiedEven the best-performing model still lags behind humans by nearly 10 percentage points.67% similarUnverifiedThe author's experience: reasoning models typically win on only 12% of tasks and lose on the economics of the remaining 88%.66% similarUnverifiedOn the strongest models, the differences between experimental groups fall within confidence intervals, and researchers characterized the results as directional patterns and soundness properties rather than proven performance lower bounds65% similarUnverified研究数据表明，在数学竞赛级别问题上，启用深度推理的模型准确率可从不足30%提升至80%以上65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1049350API
curl https://kongchang.com/api/v1/knowledge/claims/1049350MCP
get_claim(id=1049350)