正文 · AI 翻译
IT Home reported on October 27 that OpenAI announced the launch of its new model, GPT-6.1 Terra, early today (the 27th). It is designed to carry out complex computer operation tasks at lower cost and is particularly adept at work requiring visual reasoning and the use of desktop software. Compared with GPT-5.6 Terra, the new model has also improved substantially in factual accuracy and coding ability.
Terra is aimed primarily at businesses seeking to reduce automation costs and can be used for tasks such as processing invoices and interpreting technical drawings. OpenAI says Terra can also complete difficult professional tasks at a far lower cost than more expensive models.
Terra's performance improved markedly on Agents’ Last Exam, a benchmark that evaluates complex professional work. Its tasks include creating 3D models from architectural drawings and logging film shots one by one for use by film restoration teams.
At medium reasoning effort, Terra scored 15 percentage points higher than GPT-5.6 Terra while costing only about one-third as much to complete the tasks. Compared with GPT-6 Astra, Terra achieved a comparable benchmark score, while its price per token was only about one-tenth as high.
Terra also has an advantage in speed. OpenAI says it executes tasks more than 3 times as fast as GPT-6 Astra and GPT-6.1 Sol. At the highest reasoning effort, its cost per task is also only about one-ninth that of Claude Opus 5.5.
Desktop software automation is another area where Terra has made progress. OSWorld 2.1 tests a model's ability to use desktop applications through multistep tasks. At medium reasoning effort, Terra surpassed GPT-5.6 Terra's previous best score while costing considerably less per completed task.
On the MMMU-Pro visual reasoning benchmark, Terra outperformed GPT-5.6 Terra at every reasoning effort level. Despite its lower cost, its performance remained close to that of GPT-6 Astra and GPT-6.1 Sol.
OpenAI has also focused on reducing factual errors in the model to limit the effect of incorrect answers on automated workflows. The company used an internal benchmark with specially designed prompts to probe weaknesses in the model's answers.
The results showed that, at low reasoning effort, Terra's error rate was 16.7%, far below GPT-5.6 Terra's 53.6%, a decrease of 36.9 percentage points.
At the highest reasoning effort, Terra's error rate fell further to 4.3%, 18 percentage points below the previous generation and close to the levels of GPT-6 Astra and GPT-6.1 Sol.
Coding ability has also improved significantly. Terminal-Bench 4.0 uses command-line tasks to test AI coding agents' problem-solving ability, such as fixing layout shifts that occur when a webpage loads.
At the highest reasoning effort, Terra successfully completed 50.3% of the test tasks, compared with just 21.5% for GPT-5.6 Terra, giving the new model a success rate more than twice that of its predecessor. For the most difficult coding tasks, OpenAI still recommends GPT-6.1 Sol or GPT-6 Astra.
GPT-6.1 Terra became available on October 27 to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. Developers can also access it through the OpenAI API under the model identifier gpt-6.1-terra.
Under standard API pricing, 1 million input tokens cost $1 (IT Home note: about 6.7 yuan at the current exchange rate), 1 million cached input tokens cost $0.05 (about 0.33 yuan at the current exchange rate), and 1 million output tokens cost $5 (about 33.4 yuan at the current exchange rate). OpenAI says the new model costs 50% less than GPT-5.6 Terra and GPT-6.1 Sol.
On safety, Terra uses the same safeguards as GPT-6 Astra. OpenAI says the new model is more resistant to prompt injection attacks and can explain its limitations more clearly to users. Under OpenAI's Preparedness Framework assessment, Terra's cybersecurity capabilities have reached the “critical” threshold.
The lower price is expected to make complex computer automation technology easier for developers and businesses to adopt. GPT-6.1 Sol and GPT-6 Astra remain more capable of handling the most difficult coding tasks.