> Note GPT 5.5 Pro is at the top of the leaderboard only because it blew through $100 budget after only completing four cases, so 2/4 is 50%. And, a couple of other results, both Qwen models, are skewed upward in the detect % ranking because of failure to complete all cases.

Try a Wilson score interval on the lower bound of the binomial proportion confidence interval [1].

So GPT 5.5 Pro’s 2/4 (p = 0.5) for one-sided 95% (z ~ 1.645), adjusts to 0.182 [a], and the top models are revealed as the 4/9s (mimo-v2.5-pro, gpt-5.5, opus-4.8, gemini-3.5-flash and deepseek-v4). (We need to dial CI down to 76% for gpt-4.5-pro to regain top status.) If we account for speed in that cohort, derpseek-v4 (91s) is fastest followed by opus-4.8 (137s).

Given deepseek-v4 is also the cheapest model among those five, I would say—based on these data—it’s the winner. (Out of the table. If Fable got 9/9, it’s obviously first.)

[1] https://en.wikipedia.org/wiki/Binomial_proportion_confidence...