Thinking Levels
Writing Quality by Reasoning Effort
Last updated , from board data generated at 17:52 UTC.
Higher effort raised Writing Elo for most models, at very different costs. All 14 models tested at two or more explicit effort settings score a higher Writing Elo at their top setting than at their lowest. The median climb is +231 Elo, for about 4.6 times the cost per script.
The chart covers 38 models at 102 reasoning-effort settings. Each line follows one model from its lowest effort setting to its highest. Higher and further left is better: a higher score for less money or less time.
How to Use the Chart
- A hollow marker is the provider default: the level the provider applies when the request sets no effort level.
- The tabs choose the score on the vertical axis (Writing Elo, Writing Score or one of the nine metrics), and X axis switches between cost and time.
- The models on the chart are listed under it; select one to remove its line. Add or Remove Models opens the full list, up to eight at once. Reset returns to the starting models and Clear removes every line.
- Select a point for its numbers and a link to its scorecard.
On a small screen, Table lists the same numbers as rows.
Findings: Effort Against Writing Elo
At one end, Claude Sonnet 5.5 gains +766 Elo from low to max, for 22.7 times the cost. On the Elo scale, a gap that size means the top setting should score about 99% in its script comparisons against the low one, a draw counting as half. At the other end, GPT-6.1 Terra moves +6 Elo while its cost goes up 3.0 times. Their 95% Elo ranges overlap.
A model's highest setting is not always its best. For 3 of the 14 (Claude Opus 5.5, Claude Fable 5.2 and Grok 4.7), another recorded setting has a higher Writing Elo than the highest one; each model's own page below shows where it peaks. Small gaps between adjacent settings can be noise and are best read as no clear gain. The other 24 models on the chart have one explicit effort setting plus the provider default, used when the request sets no effort level, so they stay out of these numbers.
Disclosure: the Claude Opus 5.5 and GPT-6.1 Sol model lines on this chart also hold judge seats. The panel spans LLM judges from three labs plus Jev, so no lab scores itself alone, and the judge analysis measures how much each judge favors its own lab.
Each setting wrote five drafts for each of the same 10 scripts, scored blind by the same four judges as the rest of the leaderboard. All 102 effort settings on the chart sit on the leaderboard's Elo scale, since Elo comes from comparing every pair of models on every script across the full board. Cost is the price of one full script at API list prices, and time is the average wait for one accepted script, retries included. The horizontal axis is on a log scale, so equal steps mean equal multiples.
Effort Ladders by Model
Each of these 38 models has its own page with its full effort ladder, the cost of every step and a recommended setting. Each card gives the model's best setting and the model rank it earns, then the Writing Elo change from its lowest to its highest effort setting, or from the provider default when only one effort setting was run.