AI BriefingIntroducing KellyBench (2 min read)
·2026.04.10 09:00
Key point
On KellyBench, both Claude Opus 4.6 and GPT-5.4 failed to turn a profit.
Details
KellyBench simulates the 2023-24 English Premier League season to evaluate sequential decision-making in sports betting.
In testing, even models like Claude Opus 4.6 and GPT-5.4 struggled, and no model achieved a positive return.
The results show that current models have limits in accumulating experience for long-term strategy and adapting under uncertainty. This type of task requires a more complex environment.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Google revisits the hallucination problem through the lens of uncertainty
OpenAI's 'Trusted Access' Expansion Strategy for Cyber Defense and GPT-5.4 Cyber Release