Topic
Private versus public evaluations, test-set contamination, benchmaxxing and the verifiability of vendor-reported gains.
The 100% ExploitBench score and two fresh V8 zero-days are OpenAI's own numbers, one of them still unverified, but the "critical" designation is a dated document that every agent deployer's controls will now be read against.
Reality
The China Coast Guard Academy reports no weapons-rule violations from its AI command system, but the baseline it beat opens fire in 0.3% of encounters, so a few dozen simulated runs was never going to show one.
Reality
Leaked internal tests put the Exynos 2700 ahead of an unshipped Snapdragon. The number that matters is the 7.44 trillion won Samsung spent on outside application processors in six months.
Reality
Z.ai says every gain over GLM-5.2 came from post-training on broader production workflows. Whether that transfers to your stack is not something its private benchmark can tell you.
Perspective Coverage
3 publishersReality
Alibaba's Apache-2.0 Qwen3.8-27B fits in about 17GB and matched near-frontier scores, per Artificial Analysis. It also burned 3.7x the median output tokens getting there.
Reality