ToneBench eval: GPT-6.1 Terra ranks #23 at writing — cheaper, not better, than GPT-5.6 Terra
OnlyProggingForFun · reddit · 2026-10-28
Towards AI tested GPT-6.1 Terra on their ToneBench writing benchmark (10 YouTube script tasks, 5 drafts each, blind-scored by four LLM judges). Key findings:
- Cheaper, not better: 6.1 Terra ranks #23 (Elo 1923) vs GPT-5.6 Terra at #15 (2033); only 5.6 ultra clearly beats it. 6.1 on low matches most 5.6 settings for less than half the cost.
- More reasoning didn't help: zero thinking tokens at low scored the same as 9,500 tokens at max, at 3x cost.
- Judge choice matters: only one judge (GPT-6.1 Sol) preferred 6.1 Terra; the other three scored it lower.
- Strengths split: #10 of 141 configs at avoiding AI slop, but only #64 at matching the author's voice.
- Open-weight alternatives win on value: GLM-5.3 Flash max ($0.022/script) and DeepSeek V4.1 Flash 1017 max ($0.014) scored higher, though 7–15x more tokens made them slower.
Practical takeaway: 6.1 Terra low for cheap bulk drafts; GPT-6/6.1 Sol medium for quality. Caveats: narrow task scope and LLM-judge dependence.
More from Models
- GPT-6.1 Terra beats Opus 5.5 at real robot control, 3.7x better than GPT-5.6 from 3 months ago — ycombinator · 2026-10-28
- Moonshot launches Kimi K3.1 Pro: 2.8T-param open-weight MoE with 1M context, hallucinations down 65% — NVIDIAAI · 2026-10-28
- OpenSIR: a self-play framework lets LLMs teach themselves math with zero training data — PMinervini · 2026-10-28
- Claude Opus 5.5 scores 18.5% on ProgramBench, up from 0% at May launch — winer666 · 2026-10-28
- Codex background thread burns through entire $500 weekly rate limit plus 75,000 credits overnight — sethlazar · 2026-10-28
- First hands-on with Bittensor subnet SN53: open-weight models online in 2 minutes via $TAO — markjeffrey · 2026-10-28