Open-weight AI models are catching up to the frontier. The safety gap remains.
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
Open-weight models like GLM-5.2 now match frontier models in cyber and bio capabilities but refuse zero harmful requests, while closed models like Claude Opus 4.7 refuse them entirely. This means engineers shipping open-weight models must assume attackers will bypass any post-training safeguards, forcing a shift to pre-training data filtering or runtime monitoring to prevent misuse—adding cost and latency to every deployment.
Open-weight models now match frontier models in cyber/bio capabilities (GLM-5.2 performs equivalently to GPT-5.5) but lack safety guardrails, refusing 0% of harmful tasks vs. Claude's near-total refusal. This forces teams to either accept higher risk when using open models or build their own mitigations from scratch—a tradeoff that didn't exist when open models lagged in capability.