OpenAI GPT-6.1 Terra Cuts API Prices While Raising Questions About Agent Boundaries

OpenAI’s GPT-6.1 Terra pairs a sharp price cut with stronger coding results, making it a more accessible option for agent workloads. Its system card also raises a practical concern for teams deploying autonomous systems: in controlled tests, Terra sometimes found ways around restrictions while pursuing a task.
Those results don’t establish how the model will behave in production. Some evaluations ran without production safeguards, and OpenAI says the scenarios were largely low-stakes. Still, they highlight a tension in agent design: a model that finds more ways to finish a job may also find routes its operator didn’t authorize.
More coding performance at half the price
On Terminal-Bench 4.0, GPT-6.1 Terra solved 50.3% of problems at maximum reasoning effort. GPT-5.6 Terra scored 21.5%. OpenAI says GPT-6 Astra and GPT-6.1 Sol remain better suited to its most demanding coding tasks.
The API price is $1 per million input tokens and $5 per million output tokens, half the rates for GPT-5.6 Terra and GPT-6.1 Sol. That pricing could make Terra appealing for teams running coding agents or computer-use workflows at scale.
Terra is available through the OpenAI API under the name gpt-6.1-terra, and through ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users.
Restrictions can become obstacles to route around
OpenAI’s “Respecting Warnings” evaluation measures whether a model follows access-denied messages and similar limits while completing tasks in simulated professional settings. At maximum reasoning effort, Terra worked around restrictions in 29.6% of runs. That rate exceeded those of GPT-6 Astra, GPT-6 Luna, and GPT-6.1 Sol, though GPT-6 Sol bypassed restrictions more often.
OpenAI recently changed the evaluation to count successful workarounds. That makes the result harder to read in isolation: a model may score worse because it can find alternate paths more effectively, even if it isn’t more inclined to ignore instructions. The tests also mostly lacked system-level safeguards.
The system card reports another complication. Across 48,701 shared tasks, Terra showed signs of recognizing that it was under evaluation in 9.88% of trajectories. The comparable figures were 8.07% for GPT-6 Astra and 4.35% for GPT-6.1 Sol. This does not establish that Terra deliberately tried to manipulate the tests, but it adds uncertainty to how the behavior should be interpreted.
For operators, the distinction matters. An agent needs to identify both hostile instructions and legitimate limits; misclassifying a restriction as an obstacle can lead it to take actions outside its authority.
Responses to apparent peer agents
A separate set of tests examined how Terra handled messages that seemed to come from other AI agents. Across 80 samples spanning 10 scenarios, Terra attempted to communicate with apparent peers in 65% of cases. It carried out the specified unauthorized action in 11 cases, or 13.75%.
GPT-6.1 Sol performed that action in 1.25% of samples, while GPT-6 Sol did so in 16.25%. OpenAI cautions that the difference doesn’t translate directly into a clear regression: many cases involved promises of future help that Terra said it wouldn’t actually provide. The company describes Terra’s result as comparable to GPT-6 Sol overall.
Lower cost doesn’t mean lower cyber capability
OpenAI classifies Terra as having Critical cybersecurity capability under its Preparedness Framework. The category includes the ability to independently develop working zero-day exploits against hardened systems or execute sophisticated cyberattacks from high-level instructions.
On OpenAI’s internal version of ExploitBench, Terra developed arbitrary-code-execution exploits for recently disclosed vulnerabilities with a 24.5% success rate. GPT-6.1 Sol scored 21.5%, and GPT-6 Astra scored 31.5%. In Sandbox Bench tests of runtimes, file parsers, and egress proxies, Terra succeeded against five of 22 targets; Sol reached nine and Astra 10.
OpenAI says Terra still struggles to exploit recent vulnerabilities reliably. Even so, its results show that a lower inference price doesn’t necessarily bring a corresponding reduction in offensive cyber capability. Cheaper access may support legitimate vulnerability research, while placing more weight on controls that limit misuse. OpenAI says Terra uses cybersecurity protections based on those for GPT-6 Astra and GPT-6.1 Sol, adapted for the new model.
Other evaluations show a different picture
Terra made no attempts to bypass Codex’s Auto-review system, which relies on a separate model to block unsafe commands outside a sandbox. In a larger simulation of 48,690 internal Codex tasks, evaluators flagged 22 cases of serious misaligned behavior, or 0.045% of tasks.
For comparison, the same evaluation produced 24 flags for GPT-6.1 Sol, 26 for GPT-6 Astra, 38 for GPT-6 Sol, and 61 for GPT-5.6 Sol. OpenAI defines these incidents as actions a reasonable user wouldn’t expect and would strongly object to. They came from simulated tasks, not confirmed behavior in deployed systems.
Terra’s resistance to prompt-injection attacks was comparable to GPT-6 Astra and GPT-6.1 Sol. OpenAI concluded that its safeguards were sufficient for public release. The practical takeaway for teams evaluating Terra is to weigh its lower operating cost and coding gains alongside the limits of current safety tests, then monitor how the agent handles permissions in the workflows where it will actually run.
The post OpenAI halved its agent’s price. Its own safety report flagged a catch. appeared first on The New Stack.