GPT-6.1 Terra writing evaluation favors lower cost
A Towards AI evaluation tested GPT-6.1 Terra on 10 YouTube script-writing tasks, generating five drafts per task and scoring them blindly out of 100 with four LLM judges. GPT-6.1 Terra ranked 23rd overall in an evaluation of 84 models, while GPT-5.6 Terra ranked 15th. Its low-effort setting achieved overlapping Elo scores for less than half the cost of GPT-5.6 Terra at medium effort. On maximum effort, GPT-6.1 Terra used about 9,500 thinking tokens per draft and cost roughly three times as much without improving the writing score. Only one of the four judges, OpenAI's GPT-6.1 Sol, clearly preferred GPT-6.1 Terra; the other three scored it lower.
The evaluation offers a scoped comparison of cost, reasoning usage, rankings and judge preferences for a specific writing task. It does not establish general intelligence, coding or conversational ability because it covered only 10 YouTube-script tasks, used LLM judges who disagreed, and involved a particular voice-matching test.
#GPT-6.1 Terra#GPT-5.6 Terra#GPT-6.1 Sol#GPT-6 Sol#GLM-5.3 Flash#DeepSeek V4.1 Flash 1017#Towards AI