OpenAI's Terra does screen work for cents, and Morgan Stanley shows what an agent needs from your data

Terra is OpenAI's new mid-priced model, built to use software through the screen as a person does. OpenAI's models could already do this; what has changed is the price. On a test of professional software tasks, Terra scores the same as GPT-6 Astra, OpenAI's top model, at an estimated $0.29 a task instead of $4.82. For reading an invoice and entering its details into accounting software, that comes to about 20 cents.

Much routine office work happens in older software with no API, where a person copies figures from a PDF into the screen. A clerk who spends four minutes on an invoice at $30 an hour costs $2 per invoice, ten times what Terra costs.

For example, a distributor's accounts payable team enters 3,000 supplier invoices a month. Terra reads each PDF, fills in the supplier, amounts, tax and cost code, and saves the entry as an unposted draft. A clerk compares each draft with its PDF and posts it, in about a minute instead of four. The team gets 150 hours a month back for an AI bill of about $600.

The drafts are there because Terra does not get the same entry right every time, and a wrong invoice that is posted becomes a wrong payment. With drafts, the AI does the typing and a person still makes the decision that moves money. Within a year, the cheapest way to connect an old system to anything else will often be an AI that types into it.

Replit, a platform for building software, moved most of its AI spending to OpenAI after a discount of more than 30%. It could move because its product works with models from several companies, so switching meant changing a setting and running some tests.

The discount went to a customer who could show that it was able to leave. A company whose prompts and agents work with only one provider has nothing like that to show at renewal. A week of running its main AI workflow on a second provider's model gives it that proof.

Both jobs look simple, but in Morgan Stanley's wealth division each one touches several applications. The agent, which works alongside client service staff, can complete them because, years ago, the division organized its records around the client instead of each account.

Most companies keep the same customer under three numbers: one in the CRM, one in billing and one in the support desk. An agent sent to update that customer has to guess which records belong together. The dull work of giving each customer one ID decides how much an agent can be trusted to do.

The state paid A$166,000 for the study of merging three universities, and Deloitte used generative AI to help gather research and proofread it. In one table, the listed costs add up to A$659 million, but the total says A$399 million. The state has asked for its fee back.

That fee is less than a thousandth of the error, so a refund cannot cover what a wrong total costs a buyer who acts on it. A total is typed text, and no proofreading makes it match the rows above it. Add up every table yourself before the final payment.

Planning agents can weigh a stock decision across every department at once, which shows where the weeks really go. Take a delayed container of a product that sells fast. Sales wants it flown in, finance sees the airfreight bill, and logistics holds the quote. Each team is right by its own target, so the question waits for a meeting.

An agent cannot settle that either, unless someone has written down what each side is worth: the margin lost for each day without stock, and how much airfreight that justifies. Agreeing on those numbers is the real work of an agent project. Once they are written down, people can make the same call in minutes, with or without the agent.