OpenAI Unveils Cost-Effective GPT-6.1 Terra, Boosting Task Performance and Halving Costs

OpenAI unveiled its new AI model, "GPT-6.1 Terra," on the 26th (local time). Compared with the previous model, "GPT-5.6 Terra," "GPT-6.1 Terra" improves performance and cost efficiency.

OpenAI cut the standard API price for "GPT-6.1 Terra" by 50% from the previous model. The model is designed to handle everything from document writing to software operation at high speed and low cost, while supporting enterprise workflow automation.

"GPT-6.1 Terra" has strengthened communication capabilities across everyday work and also improved computer-based task performance. As a result, it is designed to be used for proposal writing, technical drawing interpretation, form entry, and video editing.

The model is also designed for automation tasks such as reading incoming invoices and entering detailed information into accounting software. OpenAI said this is intended to support automated invoice processing and accounting data entry.

OpenAI highlighted cost efficiency as a key strength of GPT-6.1 Terra and released results from the public professional-work benchmark "Agents' Last Exam." "Agents' Last Exam v1" is a benchmark that evaluates professional task performance, and OpenAI said that in this test, "GPT-6.1 Terra" scored higher than the previous GPT-5.6 Terra while lowering cost per task. In the medium-reasoning setting, GPT-6.1 Terra cost about one-third as much as GPT-5.6 Terra, while its score was 15 percentage points higher.

In the same evaluation, GPT-6.1 Terra showed performance similar to the top-tier model "GPT-6 Astra," while its per-token price was one-tenth that of "GPT-6 Astra." Based on this, OpenAI said GPT-6.1 Terra delivers higher professional-task performance at a lower cost than the previous model. The article photo was provided by OpenAI.

OpenAI said coding ability and factual accuracy also improved. Using "Terminal-Bench 4.0," which evaluates problem-solving ability in command-line environments, GPT-6.1 Terra in the maximum-reasoning setting posted a task success rate of 50.3%, while GPT-5.6 Terra's public score on "Terminal-Bench 4.0" was 21.5%. GPT-6.1 Terra's performance on "Terminal-Bench 4.0" in the maximum-reasoning setting was more than double that of GPT-5.6 Terra, and in an internal evaluation based on adversarial questions, the error rate in the low-reasoning setting was 53.6% for GPT-5.6 Terra and 16.7% for GPT-6.1 Terra, while GPT-6.1 Terra's error rate in the maximum-reasoning setting was 4.3%.

OpenAI explained GPT-6.1 Terra's performance, safety, and price together, but said GPT-6 Sol and GPT-6 Astra still have the edge on the most complex coding tasks.

GPT-6.1 Terra improved how it follows user requests, and OpenAI said it showed better adherence to the scope of requested tasks than GPT-5.6 Terra. OpenAI also said GPT-6.1 Terra strengthened the clarity of disclosures about model limits and factors that hinder task progress.

GPT-6.1 Terra showed strong resistance in prompt-injection tests. Prompt injection is a method of using malicious instructions to divert AI from the user's original task. GPT-6.1 Terra retained the safeguards applied to GPT-6 Astra. The standard API price for GPT-6.1 Terra is USD 1 per 1 million input tokens and USD 5 per 1 million output tokens.

OpenAI said GPT-6.1 Terra combines sophisticated visual reasoning, accurate computer use, and clear, natural writing. It also described GPT-6.1 Terra as a fast and efficient model that expanded the cost-versus-intelligence frontier across OpenAI's model lineup and set a new balance among speed, performance, and cost.

Source: IT DAILY · Yang Seung-gap
Original: https://www.itdaily.kr/news/articleView.html?idxno=242479

References

• https://www.itdaily.kr/news/articleView.html?idxno=242479