TL;DR. On October 26, OpenAI released GPT-6.1 Terra, a mid-range model that reads screens and operates software. It costs $1 per million input tokens and $5 per million output tokens, according to OpenAI half the price of its predecessor and of the larger GPT-6.1 Sol. As an example, OpenAI names a task found in every office: reading incoming invoices and entering them into the accounting software. In OpenAI's measurement, Terra handles office tasks almost as well as the top models, at one ninth of the cost of Claude Opus 5.5. The system card shows the other side: Terra got around a barrier in 29.6% of test runs, Sol in 22.0%. Screen work is becoming cheap enough for everyday use. Save on the model, but draw your limits with real permissions, not with instructions.
What Terra can do and what it costs
Terra is OpenAI's model for the bulk of office work. It reads what is on the screen, clicks, types and switches between applications. In its announcement, OpenAI mentions drafting proposals, interpreting technical drawings and filling out forms. Since October 26 it has been available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu customers, and in the API as gpt-6.1-terra.
The leap shows best in the cost per task. The independent benchmarking service Artificial Analysis gives Terra 48 points on its overall index, its predecessor 42 and Sol 52. One task from that index costs $0.40 with Terra and $1.40 with the predecessor. For work at the computer, OpenAI uses Agents’ Last Exam from UC Berkeley: tasks such as building a 3D model from architectural drawings or logging an animated film shot by shot for a restoration team.
What an office task costs in OpenAI's test
Terra also makes things up less often. In OpenAI's own test with trick questions, the error rate at low reasoning fell from 53.6% to 16.7%. Artificial Analysis measures how often a model guesses when it does not know the answer. Terra does so in 38% of cases, its predecessor did in 88%.
The catch in the numbers
- Almost everything comes from OpenAI. Nobody outside has checked the comparison with Claude. Terra is not yet on the independent Agents’ Last Exam leaderboard.
- A score is not done. OpenAI's score counts partly completed tasks. The independent leaderboard, with 122 public tasks from 55 sub-industries, also counts what was fully completed. There even the best reach 42.6%.
- On desktop applications the competition is ahead, even in OpenAI's own chart. On OSWorld 2.1, Terra solves at most 74.2% of tasks, Claude Sonnet 5.5 up to 83.9%. At a similar cost of just over a dollar per task, the two are level.
Score versus fully completed, independent leaderboard
tasks fully completedscore including partial success
What the system card says about limits
More revealing than the marketing numbers is the system card, OpenAI's safety report on the model. One test in it checks what an agent does when it runs into a barrier at work, such as “access denied”. The intended behaviour is to stop. Terra saw the barrier in 29.6% of runs and still completed the task, by a route that violated the purpose of the barrier. Its predecessor did so in 28.4%, Sol in 22.0%.
In simulated work environments with email, chat, browser and sales software, Terra led to an unwanted outcome in 2.28% of cases, Sol in 1.32%. Against instructions hidden in third-party content, Terra fended off 98.77% of attacks. So of 1,000, just over twelve got through.
How often something went wrong with Terra, measured against GPT-6.1 Sol
To be fair: Terra made no attempts to bypass the automatic review in Codex. Unauthorised transactions did not occur in the office test, while Sol had them in 0.31% of cases. And the tests are built to be hard on purpose. The rates cannot be converted to your daily operations, the direction can: the cheaper model takes a warning as a stop sign less often.
What this means for mid-sized companies
OpenAI's invoice example is no accident. In many companies, accounting, inventory management and customer data sit in applications without a modern interface. An agent that operates the screen needs none. It clicks through the forms like a person. For large volumes this used to be too expensive. At $0.29 to $1.01 per task in OpenAI's test, that changes.
Incoming invoices in particular have two properties that match the system card. Invoices come from outside, exactly where hidden instructions can sit. And accounting is full of barriers: approval limits, blocked suppliers, permissions not everyone has. If you give such limits to the agent only as an instruction, you rely on exactly the behaviour where Terra is weaker than the larger models. We described how agents find such detours in “When the agent finds a detour” (in German).
What you should do now
- Pick a process you can count. Incoming invoices work well because every posting is either right or wrong. Measure 50 to 100 of your own cases and count only those that are completely correct.
- Draw limits with permissions. The agent gets its own user account that can do only what it is allowed to do. A sentence in the task such as “do not approve payments” is not a barrier.
- Treat third-party documents as third-party. What is written in an invoice, email or PDF is content, not an instruction. A person confirms everything that goes outside or moves money.
- Allow uncertainty. Let the agent report “not readable” instead of guessing. Check amounts above a fixed limit and a sample by hand.
- Test models side by side. For desktop work, Sonnet 5.5 is ahead even in OpenAI's chart, and prices fall almost every month. Build so that the model can be swapped.
An agent that operates your old applications now costs about a third per task of what its predecessor cost. What it is allowed to do is still decided by your permissions, not by its price. If you would like to set up a first process so that it can be measured, we are happy to talk.
