OpenAI's GPT-6.1 Terra brings cheaper computer use to ChatGPT
TL;DR
OpenAI released GPT-6.1 Terra on 26 October. It is a fast, lower-cost AI model made to use computers, understand pictures and other visual information, and help with everyday office writing. Through the API, it costs half as much as GPT-6.1 Sol or the older GPT-5.6 Terra. OpenAI also says it makes far fewer factual mistakes. It is now available to paying ChatGPT users and through the API.
OpenAI's launch video is embedded above. You can also find it on our Videos page.
What Terra is for
OpenAI designed Terra to do computer-based tasks that people often do at work. These include writing proposals, reading technical drawings, filling in forms and editing videos. One example of automation for large amounts of work is reading incoming invoices and entering their details into accounting software.
Agents' Last Exam is OpenAI's benchmark, or test, for how well AI handles professional computer tasks. At medium reasoning effort, Terra scores 15 percentage points higher than GPT-5.6 Terra and costs about one-third as much. Terra also matches GPT-6 Astra's score while costing one-tenth as much per token. At the highest reasoning setting, OpenAI says Terra costs about nine times less per task than Anthropic's Claude Opus 5.5. These figures come from the companies that make the models, not an independent test.
For businesses, Terra's ability to avoid factual mistakes may be especially important. OpenAI tested the models using its own set of prompts, or instructions, designed to make them give wrong answers. With reasoning set to low, Terra gave wrong answers 16.7% of the time. GPT-5.6 Terra gave wrong answers 53.6% of the time. When Terra used the highest reasoning setting, its error rate was 4.3%.
OpenAI also explains Terra's limits. For the hardest coding tasks, it still recommends Astra or Sol. Under OpenAI's Preparedness Framework, Terra reaches the Critical level for cybersecurity. Because of this, Terra keeps the safety protections made for Astra.
Pricing and availability
Terra is becoming available in ChatGPT Work and Codex for subscribers with Edu, Enterprise, Pro and Plus plans. Developers can use it through the API with the name gpt-6.1-terra. Output costs $5 for every million tokens, which are small pieces of text that the model processes or produces. Input costs $1 per million tokens, or $0.05 per million tokens when cached.
Context
Terra joins GPT-6.1 Sol and GPT-6 Astra in OpenAI's current range of models. OpenAI presents Terra as the choice for tasks where speed and low cost matter most. Sol and Astra remain options for the most difficult work.
Looking Forward
For UK businesses, the example of processing invoices may matter more than the benchmark scores. In our view, a model priced for handling large amounts of routine office work could make automation cheaper for companies thinking about using AI for everyday admin. However, OpenAI's claims about factual accuracy and computer use are based on its own tests. Businesses planning to automate finance or admin tasks should test Terra on their own documents before relying on it.
Next Monday's briefing, in your inbox
Latest issue, 26 October: Mythos 5 kept attacking real targets. The week's UK AI stories, the analysis behind them, and the podcast. Free, one email a week, unsubscribe any time.
Double opt-in: we only store your address after you confirm via email. See our privacy policy.