Computer use is becoming a cost question, not only a capability question

Computer-use products do not pay only for a model's answer. They pay for every screen read, action, retry, tool call, and safety check around that answer. OpenAI's GPT-6.1 Terra launch frames the new model around that operational reality: visual reasoning and computer use at a lower cost and higher speed than the company's larger options.

OpenAI says Terra is available as gpt-6.1-terra in the API and lists standard pricing of $1 per million input tokens and $5 per million output tokens. The launch also reports improvements on internal computer-use, visual-reasoning, coding, and factuality evaluations. Those results are useful context, but they remain the vendor's measurements.

The product decision moves one layer up

The interesting question for a small team is no longer which model is strongest in the abstract. It is where the workflow needs expensive reasoning, where a faster model is enough, and where deterministic code should replace a model call entirely. A cheaper model can lower the cost of experimentation, but it does not remove the need for retries, permissions, and human review.

Model selection is becoming workflow design with a price tag attached.

Builders should measure cost per completed task rather than cost per prompt. Include failed actions, escalation to a stronger model, latency, and the human time needed to correct the result. That is the number that decides whether computer use is a product or an expensive demo.