OpenAI's GPT-6.1 Terra system card rates it Critical in cyber and High in bio/chem capability, and the safety results land on par with its predecessor rather than above it
AspireAI community · OpenAI Deployment Safety Hub
OpenAI published the Deployment Safety Hub addendum for GPT-6.1 Terra alongside the model launch, and it is a separate document from the launch post - a system card with the evaluation tables rather than the marketing summary. Under the Preparedness Framework, Terra is treated as Critical in cybersecurity and High in Biological and Chemical capability, and it ships with the same safeguards stack as GPT-6.1 Sol and GPT-6 Astra.
The safety results are mostly on-par rather than better: GPT-6.1 Terra performs on-par with GPT-5.6 Terra on Production Benchmarks, on multiturn jailbreak robustness, and on biology, chemistry and cybersecurity refusal evaluations, and it improves on GPT-5.6 Terra on U18 evaluations for age-restricted goods, disallowed sexual content, eating disorders and gore. The Safety Pareto shows improved harmful-request handling alongside helpfulness on legitimate requests. On HealthBench it scores 1.9 points below GPT-5.6 Terra after length adjustment, though it scores higher on HealthBench Professional with shorter mean responses.
The honest limit: the system card reports OpenAI's own evaluations, the HealthBench gap is a regression rather than a gain, and several table comparisons against earlier model versions may reflect later versions of those models rather than their launch-day values.
https://deploymentsafety.openai.com/gpt-6-1-terra
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.