ToneBench: DeepSeek V4.1 Flash 1017 writes better than GPT-6.1 Terra at $0.014 per script
Towards AI's ToneBench ranks GPT-6.1 Terra 23rd of 84 and concludes that for pure cheapness, GLM-5.3 Flash ($0.022) and DeepSeek V4.1 Flash 1017 ($0.014) write better for less, though they use 7-15x more tokens per script and run slower.
GPT-6.1 Terra is half the price of GPT-5.6 Terra. It doesn't write any better. We ran @OpenAI's new model on ToneBench. It's #23 of 84 models, and GPT-5.6 Terra is #15. Only 5.6 on ultra clearly beats the new model. The other three overlap in Elo range, so they're a tie, and 6.1 on low gets there for less than half the cost of 5.6 on medium. The effort setting doesn't matter either. On low, 6.1 Terra doesn't reason at all (0 thinking tokens a draft). On max it thinks about 9,500 tokens a draft and costs about 3x as much, for the same score. Only one of our four llm judges, OpenAI's own GPT-6.1 Sol, clearly prefers 6.1 Terra. The other three score it lower. OpenAI says it "writes naturally", and on clean sentences that's fair: on low it ranks #10 of 141 configs for avoiding slop. Where it falls short is sounding like me (#64 on tone and voice) and our YouTube practices (#68). I would say it's probably pretty good at conversation, less so at imitating someone's specific tone and voice. Almost like the opposite we've seen with Claude models being really "Claude-y" in a regular chat, but really good when instructed well. I'd run it on low for cheap first drafts in bulk. When the script matters, GPT-6 (and 6.1) Sol on medium beats every Terra setting, for about twice the cost of 6.1 on low. But if cheap is the whole point, two open-weight models write better for less: GLM-5.3 Flash on max ($0.022) and DeepSeek V4.1 Flash 1017 on max ($0.014). They write 7x to 15x more tokens a script, though, so they're slower. Full board: http://benchmark.towardsai.com
This post quotes another post: x.com/openai/status/2114780711664398349