OpenAI’s cheaper Terra takes on the work between apps

A cheaper route through the workday

Released on 26 October, GPT-6.1 Terra combines reading images, working across software and writing. OpenAI’s announcement gives examples including entering invoice details into accounting software and reviewing film footage. These are jobs involving both understanding the material and taking the next step on a computer.

The company says Terra completes tasks more than three times faster than Sol and Astra in its professional-software evaluation. That is a result from OpenAI’s test, rather than a promise that every campaign, spreadsheet or video edit will finish three times sooner.

What the independent tests show

Independent evaluator Artificial Analysis finds a real improvement over the previous Terra, but also a gap below Sol. At the highest reasoning setting, Terra trails Sol by four points on its broad Intelligence Index. Lower settings offer a stronger balance of score and cost. More time spent reasoning is not automatically the best bargain.

Analytics-software company Plotly supplies a more tangible example. In its 43-question exam, Terra sits between Luna and Sol. The comparison below records the cost of completing that particular exam, including the model’s intermediate work.

Plotly tested Terra through OpenRouter at its default reasoning setting, using a fabricated wind-energy dataset. It missed two deliberately hidden patterns and one question about the cause of lower turbine output. That is a useful boundary for anyone hoping an assistant will spot the important thing they did not explicitly ask about.

Does it sound like you?

Writing deserves a separate check. Louis-François Bouchard, co-founder and CTO of AI education company Towards AI, tested Terra for ToneBench, which compares scripts with his own finished YouTube work. His launch-day assessment was more positive about clean sentences than about matching his particular voice.

ToneBench uses repeated drafts and an automated judge panel, rather than a reader survey. It is a deliberately narrow task. The useful distinction for a creative team is between fluent prose and prose that sounds like the person or brand commissioning it. A general claim about natural writing does not settle that question.

Better at admitting uncertainty still leaves uncertainty

Artificial Analysis also reports that Terra is less likely than its predecessor to guess when it does not know an answer. Its reported hallucination rate fell from 88% to 38% on a private knowledge test. That measure concerns guessing when knowledge is missing; it does not mean 38% of all Terra answers are wrong.

Plotly builds analytics products that can use these models, so it has a commercial interest in the subject. Its wider methodology describes the same synthetic exam across models, with records available to researchers. The results help frame a choice; they do not establish how Terra will handle your customer data, brand judgment or a complete working day.

What access and the bill look like

OpenAI lists Terra for Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, alongside developer access. It says the same plan allowance covers more work with Terra, without specifying one universal number of tasks.

For developers, standard pricing is $1 per million input tokens and $5 per million output tokens: the units used to meter material read and generated. Cached input, where eligible repeated material is reused, costs $0.05 per million tokens. Those are usage charges, separate from a subscription. The model documentation lists additional charges for longer inputs, faster processing and regional processing; the latter adds 10% where available. We have not verified access in a Belgian account.

The interesting choice is how much judgment a task demands. Terra’s lower bill can matter for repeated, well-defined work. The independent tests leave a reason to keep a more capable model in the mix when finding an overlooked pattern matters more than shaving a few seconds off the reply.