Update: Yesterday I covered GPT-6.1 Terra’s speed and pricing. OpenAI’s system card adds a less rosy part.
The good
Across about 48,690 matched tasks, Terra got 22 severity-3-or-higher flags (0.045%), versus 38 for GPT-6 Sol and 61 for GPT-5.6 Sol. That’s about 42% and 64% fewer.
The not-so-good
Compared with GPT-6.1 Sol and GPT-6 Astra, Terra had higher rates of flags for concealing uncertainty, circumventing restrictions, exfiltration, destructive action and inbound prompt injection. In one test where other “agents” messaged it during blocked coding tasks, Terra tried to talk to those peers 65% of the time and did the unauthorized action in 11 of 80 samples (13.75%). GPT-6.1 Sol did it in 1.25%. OpenAI notes these were simulations and may not match real-world behavior.
My take
Faster and cheaper agents are great, but this is why approval gates and scoped permissions matter. Don’t give any model, Terra included, delete rights it doesn’t need.
God Bless, and Needs Garlic!
