Skip to content
NewSQL, Python and Power BI training for your team
Log in

GPT-6.1 Terra: OpenAI’s Cheaper AI That Clicks and Types for You

OpenAI’s new model sees your screen and does office tasks at half the old price. Here is what it means for Nigerian accountants and analysts.

Muhammad Jamilu Abdullahi 9 min read
Light cover reading GPT-6.1 Terra, with an amber AI cursor clicking Save on a supplier invoice form and two price labels

Picture Hauwa, an accounts officer in Kano. It's month-end, and 300 supplier invoices are waiting to be keyed into QuickBooks: supplier, date, items, VAT, total.

Now picture her typing one instruction instead. An AI looks at her screen, opens each invoice and fills in the right boxes while Hauwa checks the results.

That second picture is what OpenAI is selling with GPT-6.1 Terra, launched on 26 October. When it cut the price, one of the jobs OpenAI held up as an example was reading incoming invoices and entering them into accounting software.

For Nigerian accountants and analysts, the question isn't whether the model is clever. It's whether it's accurate, safe and cheap enough in naira to trust.

What OpenAI released

Terra is the middle child of OpenAI's four model families. They run from Luna, the cheapest, through Terra and Sol to Astra, the most capable. The new model replaces GPT-5.6 Terra.

OpenAI pitches it as a fast, efficient model that reads images in detail, operates a computer and writes in a clearer, more natural style. Its examples include drafting proposals, reading technical drawings, filling in forms and editing videos.

Plus, Pro, Enterprise and Edu users can use it in ChatGPT Work and Codex, though not in the ordinary chat window. Developers can call it through the API. GitHub made it generally available in Copilot on launch day. On X, Sam Altman said OpenAI was "especially excited for this new era of computer use."

Key factsGPT-6.1 Terra
Launched26 October 2026
API namegpt-6.1-terra
Input price$1 per million tokens (predecessor: $2)
Cached input$0.05 per million tokens
Output price$5 per million tokens (predecessor: $12)
Context window1,050,000 tokens
Maximum output128,000 tokens
ReadsText and images

OpenAI says Terra finishes computer tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol. On its own test, Agents' Last Exam, which includes jobs like turning architectural drawings into 3D models, it says Terra beats the old Terra by about 15 points at roughly a third of the cost.

OpenAI also reports 50.3% on the Terminal-Bench 4.0 coding test at maximum effort, up from 21.5%. On trick questions designed to cause mistakes, the low-effort error rate fell from 53.6% to 16.7%. All of these are OpenAI's figures.

Computer use and tokens, in plain words

Most chatbots only talk, but a computer-use model takes action. It takes a screenshot, works out what is on the screen, then clicks or types. Then it takes another screenshot to check the result and repeats until the job is done.

Think of a very fast clerk from the business centre down the road, sitting at your desk. The clerk doesn't plug into your software. They look at your screen and use it as you would.

Now imagine that clerk charges a little for every page read and more for every page written. In AI, those pages are tokens: small chunks of text, often a short word or part of a longer one. Screenshots count as tokens too.

Input tokens are what the AI reads: instructions, invoices and screenshots. Output tokens are what it writes back, including its actions. At about ₦1,370 to the dollar, Terra costs roughly ₦1,370 per million input tokens and ₦6,850 per million output tokens.

Cached input is a re-reading discount. If the same instructions are sent again and again, the repeats cost only $0.05 per million tokens. It's like a clerk who doesn't charge twice for reading the same memo.

Why it matters

The price is the headline. The input price has halved and the output price has dropped from $12 to $5. That makes repetitive screen work cheaper to automate: invoices, forms, data entry and copying figures between systems.

Computer use also works where tidy integrations don't. Many Nigerian offices rely on old desktop software and government websites with no API to connect to. An AI that uses a screen like a person could, in principle, handle them.

Some independent testing supports the efficiency claims. Artificial Analysis found that at maximum effort Terra cost about $0.40 per task, against $1.40 for the old model. On its hallucination test, Terra guessed at questions it couldn't answer 38% of the time, down from 88%.

It also lands in a price war: The Information reported OpenAI offering steep discounts to win business customers. Prices set during a fight can change.

The other side

Nearly every headline number comes from OpenAI, and Agents' Last Exam is its own test. Independent checks paint a more mixed picture.

DualEntry, a company that sells accounting software, gave Terra 101 accounting tasks and ran each one three times. Terra scored 78.2% and tied for 18th of 22 models. On accounts payable, the area closest to invoice work, it scored 53.8%. It was right in all three runs on only 65.3% of tasks.

OpenAI's safety report, the system card, is partly reassuring. Across 48,690 simulated tasks, Terra was flagged for 22 serious misbehaviours, about 0.045% and the lowest of five models tested. But it was flagged more often than GPT-6.1 Sol and GPT-6 Astra for destructive actions and for prompt injection.

Prompt injection is the biggest risk for any agent that reads screens. A web page, email or PDF can hide instructions telling the AI to drop its task and send files elsewhere. A person would ignore them; an AI might obey.

The timing is awkward. In what some called "Black October", two critical flaws were found, with AI help, in PHP, the language behind about 70% of websites. Cloudflare saw exploit attempts climb into the millions, and several European governments took websites offline as a precaution.

Banks are drawing lines. Royal Bank of Canada will block third-party personal AI agents from its online banking, calling the risk to clients "unacceptable". BMO said it doesn't let AI agents into its digital services or make transactions. JPMorgan's finance chief warned that customers who give an agent their logins may carry the blame: "that's going to be on them."

What it means for Nigeria

CBN figures put the naira at about ₦1,365 to ₦1,368 to the dollar on the official market on 26 October. To be safe, we round up to ₦1,370.

Here is a rough estimate for 1,000 invoices. The token counts are our own assumptions, so test them against your own work:

  • Reading only: the AI reads each scanned invoice and returns a table you import yourself. At 5,000 input tokens and 1,000 output tokens per invoice, that comes to $10, about ₦13,700, or ₦13.70 an invoice.
  • Full computer use: the AI opens your accounting software and keys in each invoice, taking dozens of screenshots. At 100,000 input tokens and 5,000 output tokens per invoice, that comes to $125, about ₦171,250, or roughly ₦171 an invoice.

Screen size, clicks, retries and the effort setting all change the bill. Caching helps, and OpenAI's Batch option is 50% cheaper for reading jobs that aren't urgent. In our example, clicking cost more than ten times as much as reading, so compare it honestly with what you pay staff.

Access is another hurdle. OpenAI's list of plans that get Terra leaves out the free tier and the cheaper Go plan. When OpenAI introduced naira prices in October 2025, Plus cost ₦31,500 a month, so check today's price before you pay. The API is billed in dollars, which means exchange-rate swings and card charges fall on you.

Then there is the law. Under the Nigeria Data Protection Act 2023, personal data can leave the country only under set conditions, such as adequate protection abroad or the person's informed consent. Uploading customer invoices or staff records to an AI service abroad is likely to count, so take advice first.

An advisory from NITDA's cyber-response team, reported on 5 October, warned against putting NINs, BVNs or bank details into AI chatbots. It urged organisations to adopt AI policies and to mask personal data.

Sensible first jobs include pulling invoice data into Excel or QuickBooks, filling in repetitive forms and matching bank statements to the ledger. Bank reconciliation was one of Terra's stronger areas in DualEntry's test, at 83.3%.

What you can do now

  1. Run a small pilot. Try 50 invoices and time both yourself and the AI, including the time you spend checking.
  2. Hand over the reading, not the deciding. Let the AI extract, draft and fill in. Keep approvals, payments, tax filings and client messages in human hands.
  3. Never give it banking passwords. Banks abroad are blocking AI agents and warning that customers may bear the loss.
  4. Check anything that moves money or leaves the office. Review every payment and outgoing email before it goes, even after 99 correct ones.
  5. Give it a limited login of its own. Set up a separate QuickBooks or ERP user that can enter drafts but can't approve or pay, so any mistake stays small.
  6. Mask customer data first. Strip out NINs, BVNs and account numbers, and write a short AI policy for staff.
  7. Track the bill in naira. Set a monthly budget and check usage every day during the pilot.

GPT-6.1 Terra makes an AI office assistant cheaper than ever. Treat it like a fast new clerk on probation: useful from day one, but not yet trusted with the keys.

Sources

Keep reading

All posts