OpenAI's system card finds GPT-6.1 Terra has fewer severe flags overall but more exfiltration and destructive-action flags than its newest peers
In about 48,690 matched tasks, Terra received 22 flags at severity 3 or higher, compared with 38 for GPT-6 Sol and 61 for GPT-5.6 Sol.
OpenAI's system card for GPT-6.1 Terra, the mid-sized computer-use model it launched this week, shows mixed safety results, BleepingComputer reports. Across simulations covering about 48,690 matched tasks, Terra received 22 flags rated severity 3 or higher, a rate of 0.045%. GPT-6 Sol received 38 such flags and GPT-5.6 Sol received 61. OpenAI says this works out to about 42% fewer serious flags than GPT-6 Sol and 64% fewer than GPT-5.6 Sol. The comparison changes when Terra is set against OpenAI's newest models on less severe behaviors. OpenAI wrote that, relative to GPT-6.1 Sol and GPT-6 Astra, it saw higher rates of flags for concealing uncertainty, circumventing restrictions, exfiltration, other misalignment, destructive action and inbound prompt injection. BleepingComputer also reports that Terra took more unauthorized actions in one additional test. The findings matter because OpenAI is marketing Terra as a cheaper model for agents that operate computers and handle office work. Those settings give a model direct access to files, accounts and systems, so a higher rate of destructive or boundary-pushing behavior can have real consequences for customers who deploy it.