OpenAI Cuts the Price of Working AI in Half: What GPT-6.1 Terra Means for Your Business

On Monday OpenAI released GPT-6.1 Terra, a new mid-tier model that costs half as much as the model it replaces and, on OpenAI's own numbers, roughly doubles its scores on tasks that involve actually operating a computer. API pricing starts at $1 per million input tokens and $5 per million output tokens, which also undercuts GPT-6.1 Sol, the company's stronger general model, by half.

Price cuts in AI happen so often that it is tempting to scroll past this one. We would argue you should not. Terra is not another chat model with a new name. It is aimed squarely at the unglamorous work businesses have been trying to automate for years: reading invoices, filling out forms, navigating desktop software, writing routine code.

In this post we unpack what exactly was released, what the benchmark numbers do and do not say, and how to decide whether the new model deserves a place in your product or your back office. Let's sink in!

What OpenAI Actually Released

GPT-6.1 Terra replaces GPT-5.6 Terra in OpenAI's line-up. The Terra tier sits in the middle of the family: above GPT-6 Luna, the cheap lightweight option, and below GPT-6.1 Sol and GPT-6 Astra, the models you reach for when the task is genuinely hard. OpenAI describes the new model as a fast, efficient one that combines detailed visual reasoning with precise computer use.

That positioning tells you what it is for. The company lists drafting proposals, reading technical drawings, filling out forms and editing video as typical workloads. In other words, the everyday professional tasks that are high in volume and low in glory, which is exactly where the economics of AI matter most.

Access is straightforward. Developers call it through the API as gpt-6.1-terra, and it is already available in ChatGPT Work and in Codex for Plus, Pro, Enterprise and Edu users. There is no separate waitlist, so teams can start testing the same day.

The Pricing, in Plain Numbers

The headline figures are $1 per million input tokens, $5 per million output tokens and $0.05 for cached input. That is half the price of GPT-5.6 Terra, half the price of GPT-6.1 Sol, and about one tenth of GPT-6 Astra. The cached input rate deserves a second look: agent systems resend long instructions with every step, and at $0.05 per million tokens those repeated prompts become almost free.

Per-token prices are only part of the story, though. What a business actually pays for is a completed task, and here the published numbers are more interesting. On Agents' Last Exam, a benchmark built around professional tasks, Terra scored 60.0% at $1.01 per task. Its predecessor managed 50.7% at $3.58, and the far more expensive GPT-6 Astra reached 59.3% at $7.24 per task.

For context, Anthropic's Claude Opus 5.5 still sits above all of them at 63.2%, but at $8.82 per task. If your workload needs the last few points of accuracy, you pay for them. If it does not, the gap in cost per task is now hard to ignore.

Built to Use a Computer, Not Just to Chat

The most practical change is in computer use, meaning the model's ability to operate real software: click through interfaces, fill in fields, run commands in a terminal. On OSWorld 2.1, which tests navigation of desktop applications, Terra scored 74.2% at $1.19 per task. GPT-5.6 Terra managed 57.1% and spent $4.76 per task doing it.

The jump on Terminal-Bench, a benchmark for command-line and coding tasks, is even sharper: from 21.5% to 50.3% at maximum effort. Put simply, the old model failed four of five such tasks and the new one solves about half of them, at a quarter of the token price.

Honesty requires a caveat here. Anthropic's Claude Sonnet 5.5 still leads OSWorld at 83.9%, so Terra is not the best computer-use model on the market. It is, however, by far the cheapest one that is good enough for many routine flows, and good enough and cheap is how most automation decisions in real companies actually get made.

Fewer Confident Mistakes

There is one more set of numbers that matters for anyone running AI without a human watching every step. OpenAI tested the model on prompts designed to provoke errors, the situations where models tend to invent an answer rather than admit uncertainty. At low reasoning effort, the error rate dropped from 53.6% for the old model to 16.7%. At maximum effort it fell to 4.3%, close to the company's flagship models.

Reliability is the real currency of automation. A model that is wrong half the time on tricky inputs cannot be trusted with an inbox or an order queue, whatever it costs. One that is wrong a few percent of the time can be, provided you design the process around catching those few percent.

OpenAI also reports better scope adherence and stronger resistance to prompt injection, and says the model meets the Critical cybersecurity threshold under its Preparedness Framework, keeping the safeguards developed for GPT-6 Astra. These are claims from the vendor's own report card, but they target exactly the weaknesses that have burned early adopters of AI agents this year.

The Fine Print: Where Cheaper Is Not Cheaper

Before you rush to rebuild your workflows, a few sober notes. Every benchmark figure above comes from OpenAI's own published results, and vendors choose comparisons that flatter them. Independent reviewers have already pointed out that some of the cost comparisons with competitors look selective. Treat the numbers as directionally true, not gospel.

The second point is subtler. Inference is only one line in the budget of an automated workflow. Someone still has to integrate the model, monitor it and review its output. Analyst Sanchit Vir Gogia of Greyhound Research made this point well: in his illustrative example, a handful of extra seconds of human review per task cancels out the entire inference saving. As he put it, at scale the relevant purchase is completed work that remains correct and authorised.

So the honest framing is this: Terra lowers one important cost, the per-task model spend, by several times. Whether your total cost falls depends on how much checking the output still needs, and that you only learn by testing on your own tasks.

What This Means If You Are Building Digital Products

If you run or commission software, the first practical step is to re-run your cost models. Features that were shelved a year ago because the API bill made no sense, a document-processing step, an in-app assistant, automated QA on user uploads, may suddenly clear the bar at a tenth of the projected cost. The same feature idea can flip from unprofitable to profitable purely on pricing, and this release is one of those flips.

Second, start where the volume is high and the risk is low. Invoice and document processing, form filling, support ticket triage, internal report drafting and routine coding chores are the natural first candidates. These are exactly the workloads OpenAI itself names, and they fail safely: a wrongly parsed invoice gets caught at reconciliation, not in front of a customer.

Third, measure like an operator, not like a demo audience. For each workflow, track the completion rate, the error rate, the cost per completed task and how often a human has to step in. Those four numbers tell you whether the automation pays, and they make vendor benchmarks irrelevant in the best possible way.

And finally, keep your model layer swappable. Prices and rankings in this market now change every quarter, and this week it is OpenAI undercutting everyone. Next quarter it may be Anthropic or Google. A thin abstraction between your product and the model API costs a few days of work and keeps you free to take each price cut as it comes.

Final Notes

GPT-6.1 Terra is not the smartest model you can buy, and it does not try to be. It is a bet that the next phase of AI adoption is won on price, reliability and the ability to operate ordinary software, rather than on record benchmark scores. For business owners, that bet works in your favour: the kind of automation that was a pilot project in 2025 is quickly becoming a line item any mid-sized company can afford.

Our advice stays boring on purpose. Pick one high-volume process, run the new model against it for two weeks, count completed tasks and interventions, and let those numbers, not the launch posts, decide what you roll out next. If you want help wiring AI like this into your website, your internal tools or your product, that is precisely the kind of work we do at Davydov Consulting, and we are happy to talk it through.