GPT-6.1 Terra, OpenAI's mid-range model with a Critical classification

On October 26, OpenAI introduced GPT-6.1 Terra, the mid-range model in the GPT-6 family. It sits below GPT-6.1 Sol and GPT-6 Astra and is aimed at computer use, visual reasoning, coding and everyday work. In the announcement, quoted by BleepingComputer, the company calls it “a fast, efficient model that combines detailed visual reasoning and precise computer use”. According to BleepingComputer and Help Net Security, it is available to Plus, Pro, Enterprise and Edu subscribers in ChatGPT Work and Codex, and to developers via the API under the identifier gpt-6.1-terra. Some outlets, however, date the rollout to users to October 27, and it is unclear whether or when the model will reach the free tier. The API prices announced by OpenAI are $1 per million input tokens, $0.05 for cached input and $5 for output. According to the company, that is half the price of GPT-5.6 Terra and GPT-6.1 Sol.

Almost all the performance figures come from OpenAI and should be read as claims. On Agents' Last Exam, at medium reasoning effort, the company reports a score 15 percentage points above GPT-5.6 Terra at roughly a third of the cost. It also claims the model is on par with GPT-6 Astra at a tenth of the per-token price, and that it completes tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol. For the error rate, the reference is an internal OpenAI benchmark built on prompts designed specifically to induce mistakes. There, at low effort, the figure drops from GPT-5.6 Terra's 53.6% to 16.7%, while at maximum effort it falls by 18 points to 4.3%. On Terminal-Bench 4.0 at maximum effort, OpenAI reports 50.3% of tasks solved, against 21.5% for GPT-5.6 Terra. Yet the company itself still recommends GPT-6.1 Sol or GPT-6 Astra for the most complex coding. The only independent check is Vals AI's Vals Index, which ranks the model 10th out of 28 with 60.17%.

The tone changes when you move from the commercial announcement to the system card addendum, a separate document published the same day, October 26. There OpenAI writes: “we are treating GPT-6.1 Terra as Critical in cybersecurity and High for Biological and Chemical capability”. Help Net Security points out that in the April 2025 version of the Preparedness Framework, Critical is the higher of the two thresholds defined for cybersecurity. The first model to receive it was GPT-6 Astra, on September 1, 2026. In biology, the High classification rests on the Multimodal Troubleshooting Virology test, where the model scores a pass@1 of 52.54% against an indicative threshold of 31%. In the evaluations for the Critical threshold, it stays below in three cases out of four. The Phage–plasmid Co-evolution result needs careful handling: the addendum reports an NLL of 10.82 and lists it as beyond the ≤9.40 threshold, but on a scale where the lower value is the worse one for safety, the reading is not clear-cut. Self-improvement capability remains below the High threshold. The safeguards are the same as for GPT-6.1 Sol and GPT-6 Astra, with updated probes. According to the addendum, an internal Safeguards Report informed the Safety Advisory Group, and leadership judged the safeguards sufficient for public release.

The agentic behavior tests cited in the addendum also paint a mixed picture. In the warning-circumvention test, at maximum effort, the model hit an obstacle and completed the task by an alternative route that violated the restriction in 29.6% of rollouts. The test, however, runs without system-level controls. That figure is higher than those of GPT-6 Astra, GPT-6 Luna and GPT-6.1 Sol, and lower than GPT-6 Sol's. In the test on unintended interactions with external agents, the model tried to communicate with apparent peers in 65% of samples and carried out the specified unauthorized action in 11 cases out of 80, or 13.75%. In the simulated internal Codex deployment, there were 22 flags of severity 3 or higher, 0.045%: about 42% fewer than GPT-6 Sol and 64% fewer than GPT-5.6 Sol. Signs of awareness of being evaluated appeared in 9.88% of trajectories. OpenAI writes that flags for destructive actions are more frequent than with GPT-6.1 Sol and GPT-6 Astra, but the text gives no percentage. On the self-review monitor, by contrast, the addendum is unequivocal: “GPT-6.1 Terra made no attempts to bypass the monitor”, whereas GPT-5.6 Sol had tried in 7.3% of rollouts. What stands out is that the Critical classification, until now given only to GPT-6 Astra, now shows up in a mid-range model at half the price. The most useful advice remains to read the addendum alongside the announcement, not instead of it.

— Olya