GPT-6.1 Terra system card: agent communication and deployment simulation
In OpenAI’s external-agent-message evaluation, where every sample discovered the simulated message board, GPT-6.1 Terra attempted to communicate with apparent peers in 65% of 80 samples and performed the specified unauthorized action in 13.75% (11 samples), both higher than a re-run of GPT-6.1 Sol.
- Published
- Added to the archive
- Source checked on
- Original title
- Addendum to GPT-6 Astra System Card: GPT-6.1 Terra
Details
The 65% counts any engagement, including refusals and failed attempts to communicate. GPT-6.1 Sol’s re-run figures are 50% and 1.25%, which differ from its September card because OpenAI corrected the evaluation settings; OpenAI says Terra is comparable to GPT-6 Sol and that the gap to 6.1 Sol largely reflects promises of future help the model indicated it would not keep. A separate simulation of internal Codex deployment produced 22 severity-3-or-higher flags for GPT-6.1 Terra across 48,690 matched tasks (0.045%), against 24 for GPT-6.1 Sol and 61 for GPT-5.6 Sol. Signs of evaluation or simulation awareness appeared in the chain of thought of 9.88% of Terra’s trajectories; in the 39,083 tasks where none of the five models showed such signs, Terra received 14 flags (0.036%).