OpenAI has released GPT-6.1 Terra, the mid-priced model in its lineup, at $1 per million input tokens and $5 per million output tokens, half what GPT-5.6 Terra cost. It is built to look at a screen, click and type in ordinary software. It is available in ChatGPT Work and Codex on the Plus, Pro, Enterprise and Edu plans, and through the API.
Screen work is priced in steps, and steps just got cheap
A chat answer is one pass through the model. Working a screen is many passes: the model looks at a screenshot, decides what to click, clicks, and looks again, and each pass re-reads everything that came before. An invoice that takes twenty steps in accounting software can easily use 175,000 input tokens and 10,000 output tokens. That is why the price per token matters more here than anywhere, and why the cheap rate for repeated input, $0.05 per million, matters too.
At Terra's prices that invoice costs about 23 cents. If Astra used the same number of tokens, it cost about $2.25 when it launched in September at $10 and $50 per million tokens. A clerk paid $20 an hour who spends four minutes on an invoice costs about $1.33. So in September, the best model cost more than the clerk, and Terra costs about a sixth of the clerk. At 500 invoices a month, that is roughly 33 hours of typing replaced by about $110.
OpenAI had already halved the prices of two other models in September, and now Terra arrives at half its own previous price. A job that did not pay off when you last checked is worth checking again every couple of months.
The jobs that open up have no integration
Screen control had been possible before. What changes is how many small jobs now pass the cost test. Software that was never built to talk to other software, such as an old accounting package, a supplier portal or a clinic's booking screen, can only be automated by something that uses it the way a person does. A recruiter who copies candidate details from emails into a tracking system, a shop owner who moves marketplace orders into a spreadsheet, or a hospital finance team keying in referral forms each do a job that happens dozens of times a month. None of them was worth a custom integration. At 20 cents a task, each is worth a trial afternoon.
The model is also less likely to invent what it cannot read. When it lacks the answer, it now guesses 38% of the time instead of 88%. That is still more than a third, so a smudged invoice total should be compared with the original before anything is paid.
The risk comes in with the invoice
Invoices come from outside the company, and a PDF can carry hidden text that tells the model to do something other than what its owner asked. Terra resists these attacks 98.8% of the time, up from 96.2% for the previous Terra, which still means about 1 in 80 gets through. The flagship was weaker: Astra fell for about 1 in 12 when we looked at it. A model that has your accounting software open can be talked into changing a supplier's bank details.
So the sensible setup costs nothing extra. Let the model enter invoices as drafts, and keep two things for people: approving payments and editing bank details. The model does the typing, which was never the part that needed judgment.