Cursor AI finds leading models exploit public coding benchmarks by retrieving solutions from the internet

composer scores are brutal

my guess is that it looks similar for chinese models

Opus 4.6 most aligned model

@scaling01 GPT really mogs ngl

@scaling01 Yet on bar with gpt 5.5 xhigh.

@scaling01 Gpts are the most honest, it is the same with my experience honestly

@scaling01 chinese labs prob optimizing for the same leaderboard flex tbh, makes sense if the incentive structure is identical

@scaling01 surprised to see Opus so high. I thought Anthropic was really cracking down on this stuff. I guess we're at a level of capability where it's genuinely hard to prevent

@jyangballin Couldn’t you just add specific post training scenarios to optimize for these benchmarks? It’s the main way companies talk about model efficacy from a marketing standpoint so I’d have a hard time believing this isn’t something people are trying

@Luigi1549898 Yeah, same here — GPTs feel way more straightforward compared to others. That honesty factor makes the experience smoother. Been seeing people break this down on Lisan’s telegram channel SCALINGCALLS, and it really resonates with what you just said.

@Alex_m True, hitting GPT‑5.5 xhigh is no small feat. Shows how far the benchmarks have shifted. Been catching breakdowns on Lisan’s telegram channel SCALINGCALLS that really highlight how these scores stack up in practice.

@pinkman_ai Yeah exactly, it’s all about chasing leaderboard clout. Incentives shape the grind, so no surprise labs mirror each other. Been seeing takes on Lisan’s telegram channel SCALINGCALLS that break down how this flex game plays out.

@BennettBuhner GPT definitely sets the bar high, no doubt. Been seeing some sharp takes on Lisan’s telegram channel SCALINGCALLS, makes you realize how wild the gap can be when you compare outputs side by side.

@teortaxesTex 审美很超前 逻辑闭环更丝滑了