← sources

GPT-6.1 Terra writing evaluation: lower cost, similar scores to most GPT-5.6 Terra settings

source: reddit (reddit:OpenAI)published 2026-10-28 01:11ingested 2026-10-29 02:39

retained text

We tested GPT-6.1 Terra on YouTube script writing. The main result was lower cost, rather than a clear writing-quality improvement over most GPT-5.6 Terra settings.



How we tested: 10 script tasks, five drafts per task, scored blind out of 100 by a panel of four LLM judges. The attached chart shows one entry per model using its best configuration, from the October 26 evaluation of 84 models.



What changed:



• GPT-6.1 Terra ranks #23 overall; GPT-5.6 Terra ranks #15. Only 5.6 on ultra clearly beats the new model. The other three 5.6 settings overlap in Elo range with 6.1, while 6.1 on low gets there for less than half the cost of 5.6 on medium.



• More reasoning did not improve the writing score in this test. On low, 6.1 Terra used zero thinking tokens per draft. On max, it used about 9,500 thinking tokens and cost about three times as much, for the same score.



• Judge choice matters: only one of the four judges, OpenAI's GPT-6.1 Sol, clearly preferred 6.1 Terra. The other three scored it lower.



• Clean sentences and matching a specific voice were different strengths. On low, 6.1 Terra ranks #10 of 141 configurations for avoiding slop, but #64 for matching my tone and voice and #68 for our YouTube practices.



My practical takeaway: I would use low effort for cheap first drafts in bulk. For scripts where quality matters more, GPT-6 Sol and GPT-6.1 Sol on medium beat every Terra setting in this test, for about twice the cost of 6.1 Terra on low.



Two open-weight alternatives also scored better for less: GLM-5.3 Flash on max ($0.022 per script) and DeepSeek V4.1 Flash 1017 on max ($0.014). They generated 7–15 times more tokens per script, though, so they were slower.



Scope and limitations: this measures a specific script-writing task, including how well a model imitates my voice. It is not a general intelligence, coding or conversation ranking. The scores use LLM judges, and the judge disagreement above is worth keeping in mind. My impression is that Terra may be better at natural conversation than reproducing someone's particular voice, but that is an interpretation, not something this script test establishes.



Disclosure: I'm Louis-François Bouchard, co-founder of Towards AI, and I helped run this evaluation.



document fetched 2026-10-29 00:29 · canonical https://reddit.com/r/OpenAI/comments/1xhani6/gpt61_terra_writing_evaluation_lower_cost_similar