On 13 October we asked who signs the number when ChatGPT builds the cost calculator. OpenAI has now moved the question one artefact along. GPT-6.1 Terra, released on 26 October, is a mid-tier model built for what OpenAI calls computer use: it looks at what is on the screen, clicks, types and moves between applications. The two example jobs OpenAI put in front of it were reading incoming invoices into accounting software and interpreting technical drawings.
That is a fair description of a quantity surveyor's Tuesday. The price is what OpenAI wants you to notice. Terra costs $1 per million input tokens and $5 per million output, half the rate of GPT-6.1 Sol and half what the previous Terra cost. It is live in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu subscribers, and in the API. A model this cheap, marketed at invoices, drawings and forms, makes the checking regime the thing worth discussing, ahead of the price.
What OpenAI is selling
OpenAI's launch post, as relayed by The Register, says the price makes it practical to automate high-volume work such as reading incoming invoices and entering their details into accounting software. Ted Sanders, a member of technical staff at OpenAI, was more measured in a Hacker News thread the same day.
Terra is probably better than Astra and Sol at technical communication, similar on vision and computer use (but much faster!), and a bit worse on most other things.
He added that for subscribers "it's not a bombshell release". We read that as the honest version of the pitch. Terra reads a drawing about as well as the top models, much faster, and at a tenth of Astra's token price.
Sam Altman's launch post was shorter still.
6.1 Terra is here! It is good at a lot of things, but we are especially excited for this new era of computer use.
Computer use means the model does the clicking, which is where a misread dimension becomes a wrong entry in a live system.
The numbers and what they measure
- Hallucination rates: OpenAI reports a fall from 53.6% to 16.7%, comparing GPT-5.6 Terra with the new model at low reasoning effort. At maximum effort, the new model records 4.3%.
- What those figures mean: The evaluation uses ChatGPT conversations flagged by users for factual errors, deliberately selecting cases prone to hallucinations. OpenAI cautions that these results do not reflect hallucination rates in everyday use. The improvement is encouraging, but it does not establish how reliably the model performs in production.
- Independent assessment: Artificial Analysis found a similar improvement in its AA-Omniscience benchmark. The new Terra guesses when it does not know an answer 38% of the time, down from 88% for its predecessor. That is a substantial improvement, but the model still gets more than a third of these situations wrong.
- Speed and performance: OpenAI claims Terra completes tasks more than three times faster than Sol and Astra on its Agents' Last Exam benchmark. NerdsChalk's smaller test found a more modest improvement: Terra completed a supplier form in 9.4 seconds, compared with 17.3 seconds for Sol, while both models answered every form field and drawing question correctly across all runs.
- How much weight to give the results: NerdsChalk tested just three tasks, three times each, on one machine over a single night. The findings offer a useful comparison but are not enough to establish a consistent performance advantage.
The row worth staring at is Luna, OpenAI's cheapest model. It returned an empty form in two of three attempts and once missed both calculated answers on the drawing. The Neuron's Eric Gerard Ruiz and Corey Noles drew the conclusion we would: "Vendor numbers are averages across their tasks, and your task is the one that matters."
From the calculator to the drawing
Our 13 October piece examined the RICS professional standard on AI, which came into force in March. It requires a named surveyor to document why an AI output that materially affects their service can be trusted.
The same requirement applies when a model reads a general arrangement drawing to prepare a measurement schedule or enters a subcontractor's invoice into the ledger. The difference is that these tasks can happen with far less human scrutiny, often without anyone stopping to question the output. The model just does the task.
Two features of Terra sharpen this. OpenAI rates it ‘Critical’ for cybersecurity under its Preparedness Framework, the same as GPT-6 Astra, and ships it with the same safeguards. The system card reports strong resistance to indirect prompt injection, where instructions hidden in a web page or a document try to divert the agent, and an agent reading supplier PDFs is reading content you did not write. And it is cheap enough that the natural failure mode is volume. One wrong invoice is a query. Four hundred processed overnight is a reconciliation exercise.
Takeaway
Terra is the first model we have seen priced and marketed for the routine artefacts our readers handle, and the early independent evidence says it matches the dearer models on short, well defined tasks. It does not change who owns the number. Before it touches live project data:
Run the race The Neuron describes your own documents: three runs each of Terra and a dearer model on a form or drawing you can mark, with dummy data, and switch only if accuracy holds.
Decide in writing what a human checks on every output and what gets sampled, because the RICS standard asks for that document rather than the benchmark chart.
Keep agent runs on supplier documents in a sandbox until you have watched it handle content that was never written for a machine.
Read the 16.7% as OpenAI reads it: a result on a hard test set, with production rates unknown.
The question from 13 October has moved from the calculator to the drawing, and it is now a tenth of the price to get wrong.
Every week we read the model launches so you can get back to the drawings. The Project Flux newsletter puts the price, the caveat and the checking question in one email at projectflux.ai.
Links and Stuff
All content reflects our personal views and is not intended as professional advice or to represent any organisation.
