Model Directory
AI Model Writing Scorecards
Last updated , from board data generated at 20:08 UTC.
Scorecards for the 80 models on ToneBench, one page for each of their 132 configurations. A configuration is one model at one reasoning setting. The grid opens on Best per model: each of the 80 models once, at its best configuration (the setting with the highest Writing Elo), in model rank order. All configurations shows every setting with its configuration rank. Ranks come from Writing Elo, so a card can show a higher Writing Score than the card above it.
Key Results
No other model's best configuration has a 95% range that overlaps it.
All 6 models above it are proprietary; its best configuration is 576 Elo behind the #1 model's.
The highest-ranked model whose best configuration costs $0.052 per script or less, the cheaper half of the models.
The same model's configurations rank this far apart depending on the reasoning setting.
Computed from the current board. These results change as models, reference scripts or scoring criteria are added.
- Each card shows a rank, the Writing Score (the 0 to 100 consensus of the four judges and code), the Writing Elo (a rating from pairwise comparisons of those scores) and the cost of one script at API list price (a route that served a model free shows $0). Under Best per model the rank is the model rank, marked model; under All configurations it is the configuration rank, marked config.
- Writing Elo is a relative scale: two configurations sit below zero because they lost far more script-by-script comparisons than they won.
- Search by model, lab or setting name, show only open-weight or proprietary models, or sort by rank, Writing Score or cost with the controls below. A search covers all configurations in either view and shows each match with its configuration rank.
- Select a card to open its full scorecard: nine metrics, results on each of the 10 reference scripts, consistency, cost and the evaluation setup. Every ranked scorecard also has a menu to compare that model with any other.
All 132 configurations completed all 10 reference scripts and hold a current rank.
Showing all 80 models, each at its best configuration.
* Estimated from API list prices. Free route: the route used served the model at no charge.
No model or setting matches this search. A search covers every configuration, under Best per model and under All configurations. Clear the search box to see every model at its best configuration.
Frequently Asked Questions
How Many AI Models Does ToneBench Test for Writing?
132 configurations of 80 models are on the board, from 15 labs. A configuration is one model at one reasoning setting, so the same model at low and at max effort counts twice; the setting alone can move a model a long way. Proprietary models from Anthropic, OpenAI, Google and xAI sit next to open-weight models such as DeepSeek, Qwen, Kimi and GLM, so a model you can run yourself is measured on the same work as the proprietary ones. Every ranked configuration gets the same 10 briefs, each written for a What's AI video, and four judges score every draft blind, with the channel's finished script as the reference.
Which Model Writes Best Overall?
Claude Opus 5.5 ranks #1 of 80 models, each model ranked by its best configuration. Its best configuration, Claude Opus 5.5 (xhigh), has a Writing Elo of 2785 and a Writing Score of 85.8/100. It costs $1.00 per script.
What Is the Best Open-Weight Model for Writing?
GLM-5.3 is the highest-ranked open-weight model, #7 of 80 models, through its best configuration, GLM-5.3 (max): configuration #15 of 132, Writing Elo 2209, at $0.240 per script. That is 576 Elo behind the #1 model, a gap at which the #1 model is expected to score about 96% in script-by-script comparisons, a draw counting as half, so the strongest proprietary configurations still rank above it on these scripts. It is the first one to try when you need to run the model yourself or keep data in-house.
Which Model Writes Best in the Cheaper Half of the Board?
With each model priced at its best configuration, the cheaper half of the models costs $0.052 per script or less. Among them, GLM-5.3 Flash ranks highest, #12 of 80 models, through GLM-5.3 Flash (max): Writing Elo 2126 at $0.022 per script. The #1 model costs more than 40 times as much per script, for 659 more Elo.
How Many of the Models Have Open Weights?
41 of the 132 tracked configurations have open weights, so they can be downloaded and run locally. The rest are proprietary. The Open weights and Proprietary filters above the grid show each group on its own.
How Often Is This List Updated?
New models are added soon after release, and new reference scripts join the set from time to time. Each new script is normally benchmarked before its video is published, which lowers but does not remove the chance that a model has seen it. Every card and score is rebuilt from the current board, and the methodology explains the scoring.
How Do I See Detailed Scores for a Specific Model?
Search for it by model or lab name, then select its card. The scorecard leads with the model rank and the configuration rank, then lists all 9 metrics with their rank among the configurations and the result on each of the 10 reference scripts, which shows whether a configuration leads everywhere or only on the scripts that suit it.