ToneBench
Towards AI · 2026-08-14-10-task · Blind, panel-scored writing benchmark across a fixed set of real editorial scripts.
Model Ranking
Source | |||||||
|---|---|---|---|---|---|---|---|
| #1 | Claude Opus 5Anthropic | 90.37 | 2,832 | 50 | Claude Opus 5 (max effort)average per case | Original sourcewinner eligible | Third-party benchmark |
| #2 | Claude Fable 5Anthropic | 90.35 | 14,132 | 50 | Claude Fable 5 (max effort + 4.8 fallback)average per case | Original sourcewinner eligible | Third-party benchmark |
| #3 | Kimi-K3Moonshot AI | 89.26 | 15,418 | 50 | Kimi K3average per case | Original sourcewinner eligible | Third-party benchmark |
| #4 | GPT-5.6 SolOpenAI | 88.47 | 3,470 | 50 | GPT-5.6 Sol (high)average per case | Original sourcewinner eligible | Third-party benchmark |
| #5 | Grok 4.6xAI | 88.14 | 15,288 | 50 | Grok 4.6 (high)average per case | Original sourcewinner eligible | Third-party benchmark |
| #6 | Claude Opus 4.8Anthropic | 88.09 | 13,476 | 50 | Claude Opus 4.8 (max effort)average per case | Original sourcewinner eligible | Third-party benchmark |
| #7 | GPT-5.6 TerraOpenAI | 86.97 | 2,688 | 50 | GPT-5.6 Terra (ultra)average per case | Original sourcewinner eligible | Third-party benchmark |
| #8 | GPT-5.6 LunaOpenAI | 84.80 | 2,962 | 50 | GPT-5.6 Luna (ultra)average per case | Original sourcewinner eligible | Third-party benchmark |
| #9 | Qwen3.8-MaxQwen | 83.35 | 20,929 | 50 | Qwen3.8 Maxaverage per case | Original sourcewinner eligible | Third-party benchmark |
| #10 | Gemini 3.1 ProGoogle DeepMind | 80.16 | 8,033 | 50 | Gemini 3.1 Pro (default)average per case | Original sourcewinner eligible | Third-party benchmark |
| #11 | Gemini 3.7 FlashGoogle DeepMind | 78.90 | 5,270 | 50 | Gemini 3.7 Flash (default)average per case | Original sourcewinner eligible | Third-party benchmark |
| #12 | Gemini 2.5 ProGoogle DeepMind | 78.63 | 4,636 | 50 | Gemini 2.5 Proaverage per case | Original sourcewinner eligible | Third-party benchmark |
| #13 | Gemini 3.6 FlashGoogle DeepMindSelected model | 78.39 | 9,421 | 50 | Gemini 3.6 Flash (high thinking)average per case | Original sourcewinner eligible | Third-party benchmark |
| #14 | Gemini 3.5 FlashGoogle DeepMind | 75.70 | 8,870 | 50 | Gemini 3.5 Flash (default)average per case | Original sourcewinner eligible | Third-party benchmark |
| #15 | Gemini 3.5 Flash-LiteGoogle DeepMind | 74.30 | 2,022 | 50 | Gemini 3.5 Flash-Liteaverage per case | Original sourcewinner eligible | Third-party benchmark |
| #16 | Gemini 2.5 FlashGoogle DeepMind | 73.83 | 4,339 | 50 | Gemini 2.5 Flashaverage per case | Original sourcewinner eligible | Third-party benchmark |
| #17 | Mistral Large 3Mistral AI | 70.61 | 2,014 | 50 | Mistral Large 3average per case | Original sourcewinner eligible | Third-party benchmark |
| #18 | Gemini 3.1 Flash-LiteGoogle DeepMind | 69.50 | 1,571 | 50 | Gemini 3.1 Flash-Liteaverage per case | Original sourcewinner eligible | Third-party benchmark |
| #19 | Gemini 2.5 Flash-LiteGoogle DeepMind | 68.49 | 2,405 | 50 | Gemini 2.5 Flash-Liteaverage per case | Original sourcewinner eligible | Third-party benchmark |
| #20 | GPT-4oOpenAI | 57.17 | 1,184 | 50 | GPT-4oaverage per case | Original sourcewinner eligible | Third-party benchmark |
| #21 | Command ACohere | 55.67 | 1,148 | 50 | Cohere Command A (03-2025)average per case | Original sourcewinner eligible | Third-party benchmark |
| #22 | GPT-4o MiniOpenAI | 54.96 | 1,394 | 50 | GPT-4o miniaverage per case | Original sourcewinner eligible | Third-party benchmark |
22 models ranked by the best winner-eligible current-version score when available, otherwise the best publisher-artifact score; highest first. Observed Aug 29, 2026. The model you came from is highlighted.
Methodology and Coverage
Primary Evidence
Publisher Artifacts
Questions
ToneBench FAQs
What does ToneBench measure?+
Blind, panel-scored writing benchmark across a fixed set of real editorial scripts. Model Markets classifies it as a writing benchmark and preserves the publisher's 2026-08-14-10-task release as a distinct comparison cohort.
How are models ranked on ToneBench?+
Models are ordered by overall score in points, with higher scores ranked first. Each model appears once; a verified winner-eligible current-version result takes precedence over a publisher-artifact-only result, and missing scores are not estimated.
Which model currently leads ToneBench?+
Claude Opus 5 leads the current verified table with 90.37 points on 2026-08-14-10-task. This is a benchmark-specific result, not a universal model-quality claim.
How many models have a published ToneBench score?+
22 catalog models are published from 136 source rows. 81 source identities remain quarantined rather than guessed.
Can ToneBench scores be compared with other benchmarks?+
Raw scores should be compared only within the same benchmark version, metric, and protocol. Model Markets normalizes eligible scores only for aggregate rankings, and groups models by an identical benchmark set before ranking them. View aggregate rankings →
Why might a model be missing from ToneBench?+
A model remains absent when the publisher has no current result, the source model identity is unresolved, the evaluation protocol is incompatible, or the evidence cannot be verified. Model Markets does not substitute a provider claim or infer a score from a related model.
How current is the ToneBench leaderboard?+
The current Model Markets snapshot was observed Aug 29, 2026 from Towards AI artifacts. The exact publisher source and each retained result artifact are linked on this page.