Model Scorecard
Qwen3 235B Thinking (thinking always on)
Alibaba (Qwen) · open weights
Qwen3 235B Thinking ranks #70 of 87 models on the November 2026 ToneBench board, and Qwen3 235B Thinking (thinking always on) is its only configuration, #131 of 151 configurations, with a Writing Score of 62.6/100 and a Writing Elo of 602. One script costs $0.012 at API list price.
Last updated , from board data generated at 17:52 UTC.
Rank, scores, the nine metrics, the 10 reference scripts, cost and setup for one configuration (a model at one reasoning setting). Key Results sums it up; the section links below it jump to the detail.
Key Results
- StandingQwen3 235B Thinking ranks #70 of 87 models, and this is its only configuration on the board. As a configuration, it is #131 of 151 configurations by Writing Elo (602), in the bottom third of the board. Its 95% range (559 to 644) overlaps 8 other configurations, ranked #130 to #138.
- MetricsBest-ranked: Length Adherence (79.8, 89th of 151 configurations, -5.0 against the board median). Lowest-ranked: Slop Score (57.1, 146th of 151 configurations, -24.2 against the board median).
- ScriptsFrom 59.9 on Scripts 4 and 10 (short explainer and engineering process walkthrough) to 67.8 on Scripts 2 and 8 (news-analysis and product release announcement): a 7.9-point spread across the 10 reference scripts.
- ConsistencyIts Writing Score moves ± 3.7 between drafts of the same script, against ± 2.5 for the median configuration: more variable than most.
- ValueGPT-6 Luna (medium) ranks far higher and costs about $0.0026 per script, against $0.012 per script for this one.
Louis's PickI'd pick DeepSeek V4.1 Flash 1017 (high) over this one: it ranks clearly higher for no more per script.
More Context Behind These Results
Writing Elo comes from comparing its average on each script with every other configuration's average on that same script; a gap smaller than the draft-to-draft variation counts as a draw (what the ranges can and cannot show).
Slop Score ranks in the bottom half (146th of 151 configurations), so expect filler phrases and AI tells to clean out.
The script format moves its score a lot. If most of your scripts look like its best or weakest format, that result is a better guide than the average.
It outranks 4 configurations that cost at least 50% more.
Among open-weight models, Qwen3 235B Thinking is 21st of 34. As a configuration, it is 28th of 43 open-weight configurations. The best open-weight configuration is GLM-5.3 (max), at configuration #22, and every configuration above it is proprietary.
Skill Profile Across the Nine Metrics
Each corner is one metric on a 0 to 100 scale; the further out, the higher the score. A corner outside the dashed line beats the board median. Hover or tap a point for its exact score, and a metric name for what it rewards.
Scores by Metric
Nine metrics, each scored 0 to 100 and weighted into the Writing Score; the weight sits under each name. Rank is out of all 151 ranked configurations.
Scores by Reference Script
Every ranked configuration writes five drafts for each of the same 10 briefs, and each score is the mean of those drafts. The rank and board median compare it with every configuration on that script, which separates a weak result from a hard script. The result on the format closest to your own work is the most relevant.
Consistency Between Drafts
Every score on this page is the mean of 5 drafts per script. The ± is the standard deviation between those drafts: the larger it is, the more two drafts of the same script can differ.
Compared with the typical model, it varies most on Substance, Accuracy & Value: ± 4.7 between drafts against ± 2.1, so two drafts of the same script can score quite differently there.
Consistency by Metric
| Metric | This model | Board median | Verdict |
|---|---|---|---|
| Tone & Voice Match ? | 54.4 ± 3.2 | ± 1.9 | more variable than most |
| Writing Craft & Clarity ? | 62.5 ± 2.2 | ± 1.5 | more variable than most |
| Substance, Accuracy & Value ? | 59.0 ± 4.7 | ± 2.1 | more variable than most |
| Continuity & Emotion ? | 64.3 ± 2.7 | ± 1.5 | more variable than most |
| YouTube Best Practices ? | 63.8 ± 5.4 | ± 2.4 | more variable than most |
| Hook Strength ? | 72.2 ± 3.3 | ± 2.7 | typical |
| Length Adherence ? | 79.8 ± 10.2 | ± 6.6 | more variable than most |
| Slop Score ? | 57.1 ± 2.2 | ± 1.6 | typical |
| Visual Cue Quality ? | 56.0 ± 6.0 | ± 2.9 | more variable than most |
Cost, Time and Tokens per Script
Logged on every run; none of these figures affects the Writing Score. Time is the average time to one accepted script, retries included. Cost is one script at the API list price.
Scores by Judge
Four judges score the same 50 stored drafts blind against the same rubric: three LLM judges from three labs, plus Jev, a typed judge from TypeSafe that rates each metric on five described quality levels. No single judge's taste decides the result. The Writing Score of 62.6 combines their scores with the parts computed in code (Length Adherence and half of the Slop Score). The highest and lowest judge are 20.7 points apart on its Writing Score. That is wider than the board median of 13.4: GLM 5.3 Flash rates it noticeably higher than Claude Opus 5.5 (medium), so the Writing Score sits between two different opinions. Each judge also has its own scale, so every card shows that judge's score against its average across the board. How the panel works.
Evaluation Setup
How this exact configuration was run and judged: route, model ID, dates, sampling and prompt.
- Accessed viaOpenRouter · model page ↗.
- Exact model / route id
qwen/qwen3-235b-a22b-thinking-2507(the request sets no level; the provider applies thinking always on). - Generations5 independent runs × 10 scripts = 50 drafts, stored after generation and never edited.
- Run dates2026-10-08.
- JudgesClaude Opus 5.5 at medium effort through the Claude Code subscription, GPT-6.1 Sol at medium reasoning through the Codex CLI, GLM 5.3 Flash through OpenRouter, and Jev 1.13 from TypeSafe. One call per judge per draft. All four score every draft on their own, blind to which model wrote it. The three LLM judges use the identical rubric and have to quote evidence before they score a metric. Jev answers the same rubric as one question per metric, choosing among five described quality levels, and writes no text. Seven metrics are the plain average of the four judges. Length is computed from the spoken word count, and Slop is half computed (lexical scans minus the spoken-dash penalty) and half judged. The Writing Elo is cross-checked with a Bradley-Terry fit of the same pairwise results (methodology).
- Sampling and promptThe provider's default sampling, with no fine-tuning and no system-prompt changes. Every current-ranked model gets the same prompt, about 13,000 tokens (the count shifts a bit with each model's tokenizer): a style guide built from Louis's voice and storytelling guides, the target length, and the brief with its research packet. Generation is closed-book on every route: the prompt is the only source, with no web search, browsing, code execution, file access or other tool (generation).
- What stays privateThe exact rubric and any non-public reference material. Each script is normally benchmarked before its video is published, which lowers the odds that a model has seen the reference, though it does not rule it out (why).
Frequently Asked Questions
How Good Is Qwen3 235B Thinking (thinking always on) at Writing?
Qwen3 235B Thinking ranks #70 of 87 models on ToneBench, and this is its only configuration on the board. This configuration is #131 of 151 configurations, in the bottom third of the board, with a Writing Elo of 602 and a Writing Score of 62.6/100. The ranks come from full YouTube scripts written in the voice of Louis-François Bouchard and judged blind, with the channel's finished script as the reference. The benchmark measures long-form scripted writing, so the result does not carry over to every kind of text.
Is Qwen3 235B Thinking (thinking always on) the Best Alibaba (Qwen) Model for Writing?
No. Qwen3 235B Thinking is 4th of 8 Alibaba (Qwen) models. The top one is Qwen3.8 Max, through Qwen3.8 Max (thinking on), about 930 Elo ahead of this setting.
Is Qwen3 235B Thinking (thinking always on) Good Value for the Money?
It costs about $0.012 per script. GPT-6 Luna (medium) ranks higher and costs less.
What Are the Strengths and Weaknesses of Qwen3 235B Thinking (thinking always on)?
Its best-ranked metric is Length Adherence (79.8, 89th of 151 configurations, -5.0 against the board median), and its lowest-ranked is Slop Score (57.1, 146th of 151, -24.2). Its best-ranked metric still sits in the bottom half. The Scores by Metric section shows the board median and the board's best for each one.
How Was Qwen3 235B Thinking (thinking always on) Evaluated?
It was run through OpenRouter, with the exact model/route ID qwen/qwen3-235b-a22b-thinking-2507, first run on 2026-10-08. It wrote each script 5 times with the provider's default sampling and no fine-tuning. The Evaluation Setup section and the methodology page have the details.
Does Qwen3 235B Thinking (thinking always on) Have Open Weights?
Yes. Its weights are public, so it can be downloaded and run on your own hardware. Check the license if you need open source in the strict sense. Among open-weight configurations it is 28th of 43 for writing.