GPT-6.1 Terra runs desktop tasks at half the API price of GPT-5.6 Terra
OpenAI's new fast, low-cost model operates software on screen and costs 50% less per token than GPT-5.6 Terra and GPT-6.1 Sol. For back-office teams, the cost of automating screen-based work just dropped.
OpenAI has released GPT-6.1 Terra, the new version of its fast, lower-cost model. It reads images and screens in detail, operates software the way a person would, and writes more naturally than GPT-5.6 Terra, the model it succeeds. Standard API prices are 50% lower than both GPT-5.6 Terra and the larger GPT-6.1 Sol: $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. It is live now for Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and in the API as gpt-6.1-terra.
A cheaper model that works a desktop
The main claim is about computer use: a model that looks at a screen and then clicks and types its way through a task. OpenAI tested this on Agents' Last Exam, a benchmark of professional tasks such as turning architectural drawings into a 3D model or logging an animated film shot by shot. At medium reasoning effort, Terra scored 15 percentage points higher than GPT-5.6 Terra at about one-third of the cost. It matched GPT-6 Astra, OpenAI's most capable model, at one-tenth of Astra's price per token. OpenAI says it also finishes tasks over 3 times faster than Astra or Sol.
At maximum effort, OpenAI puts Terra's cost per task at about 9x lower than Anthropic's Claude Opus 5.5: it uses about a third of the tokens, at a quarter of the price per token.
Fewer wrong answers, better code
The second change is accuracy. On OpenAI's internal set of prompts built to provoke mistakes, the error rate at low reasoning effort fell from 53.6% with GPT-5.6 Terra to 16.7%. At maximum effort it fell to 4.3%, close to Astra and Sol.
Coding improved too. On Terminal-Bench 4.0, a test of coding agents working from the command line, Terra solved 50.3% of tasks at maximum effort. GPT-5.6 Terra's published score was 21.5%. OpenAI still points to Sol and Astra for the hardest coding work.
All of these are OpenAI's own figures, on benchmarks it chose, so run your own test before switching anything.
What it means for back-office work
Terra is aimed at the work that fills a shared-services floor. OpenAI's own example is reading incoming invoices and entering their details into accounting software. Filling in forms and checking documents against records are the same kind of job. In many companies that work runs through screens with no API, so automating it meant custom integration or a robotic process automation script that breaks when a screen changes. A model that reads the screen and acts on it is a different route, and at half GPT-5.6 Terra's price the cost case for trying it has changed.
Price matters most at volume. A team processing thousands of invoices a month pays for every token the model reads. Cached input costs $0.05 per million tokens, against $1 for normal input, which helps when the same instructions and templates repeat on every document.
What to do next
- Pick 1 repetitive screen task. Invoice entry and claims intake are good candidates. Run it on Terra in a sandbox with real samples and compare its error rate with your current process.
- Limit what the agent can touch. Computer-use agents act on live systems. Start with test accounts and read-only access. OpenAI rates Terra at its "Critical" level for cybersecurity capability under its Preparedness Framework. It keeps the safeguards built for GPT-6 Astra, and OpenAI reports strong resistance to prompt injection in its tests.
- Reopen stalled business cases. If an automation pilot died on cost in the past year, rerun the numbers at the new prices.