GPT-6.1 Terra makes task routing the production default
OpenAI’s new model changes the economics of visual and desktop workflows, but it justifies a new routing lane rather than a full-stack migration.
What launched
OpenAI released GPT-6.1 Terra on October 26 for visual reasoning, application control and writing. It is available through the API, ChatGPT Work and Codex, with a 1.05-million-token context window and standard pricing of $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens.
Tool calling requires the Responses API. Requests above 272,000 input tokens carry higher rates, so the headline price is not enough to forecast a production bill.
Economics, not parity
Terra is not a cheaper copy of the strongest model. On Plotly’s data-analysis test it completed 40 of 43 tasks, while Sol completed all 43, but Terra’s full run cost half as much and recorded a lower median time per question.
Artificial Analysis scores it at 48 and places it among the stronger models in its price range, while also recording higher output-token use than the median. Its advantage therefore depends on the whole task, not the rate card.
What we would do
We would now place bounded, repeatable and easily verified workflows into a Terra shadow lane: invoice extraction, document classification, visual matching and controlled actions inside applications. Complex coding and irreversible operations would stay on a stronger model.
Migration decisions should follow cost per accepted completion, total latency, escalation rate and recovery from failure. The architecture should retain a fallback and avoid binding the entire system to one provider.
GPT-6.1 Terra is not the model we would put everywhere; it is the model that makes a production stack without a cheaper execution lane increasingly difficult to justify. Its standard price of $1 per million input tokens and $5 per million output tokens makes it a serious candidate for document processing, visual matching, form completion and bounded actions in desktop applications. Plotly recorded 40 correct results out of 43, against Sol’s 43, at half the test-run cost and with a lower median time per question. That is enough for us to place Terra into shadow production now, but not enough for a blind migration. OpenAI still recommends Sol or Astra for the most complex coding tasks, while independent results describe better economics rather than universal superiority. We would route bounded, verifiable work to Terra, measure cost per accepted completion and end-to-end latency, and escalate complex, risky or irreversible actions to a stronger fallback. Teams waiting for one model to win every benchmark are postponing an optimization they can test safely today.
Turn insight into a working system.
We build a shadow-routing and evaluation layer that sends bounded visual and desktop tasks to Terra while keeping harder or irreversible work on stronger fallbacks. We add permissions, audit trails and cost-per-completion reporting. Talk to us if your team runs document or computer-use workflows in production.