Sign inTest your prompt

Leaderboard / Spring Prompt Score

Measured by Spring Prompt

Spring Prompt Score

One number per model from our own benchmarks only: CatalogBench, DeckBench, SurveyBench, ROASBench and BulletBench. Every benchmark counts the same (what each one tests). Because it uses no third-party results, you are free to cite it, download it and embed it.

Models scored
28
Benchmarks
5, all ours
Results up to
10 Nov 2026
Score
0–100, higher is better
Spring Prompt Score from 0 to 100, higher is better, with how many of our benchmarks it covers and the model's percentile on each
#ModelScoreCoveragePercentile on each benchmark
1= Claude Fable 5.2Anthropic 90 5 of 5 benchmarks CatalogBench 85DeckBench 93SurveyBench 100ROASBench 100BulletBench 74
1= GPT-6.1 SolOpenAI 90 4 of 5 benchmarks CatalogBench 95DeckBench 89SurveyBench 82ROASBench –BulletBench 92
3= GPT-6 SolOpenAI 84 4 of 5 benchmarks CatalogBench 90DeckBench 96SurveyBench –ROASBench 86BulletBench 64
3= GPT-6 AstraOpenAI 84 4 of 5 benchmarks CatalogBench 99DeckBench 100SurveyBench –ROASBench 93BulletBench 43
5 Claude Opus 5.5Anthropic 79 4 of 5 benchmarks CatalogBench 66DeckBench 86SurveyBench 88ROASBench 75BulletBench –
6 GPT-6.1 TerraOpenAI 77 5 of 5 benchmarks CatalogBench 87DeckBench 79SurveyBench 71ROASBench 96BulletBench 53
7 Muse WonderMeta 67 5 of 5 benchmarks CatalogBench 85DeckBench 57SurveyBench 94ROASBench 39BulletBench 57
8 GPT-6 LunaOpenAI 65 4 of 5 benchmarks CatalogBench 76DeckBench 68SurveyBench –ROASBench 82BulletBench 33
9 Claude Fable 5.1Anthropic 60 3 of 5 benchmarks CatalogBench 34DeckBench 75SurveyBench –ROASBench 71BulletBench –
10 Grok 4.7xAI 56 3 of 5 benchmarks CatalogBench 72DeckBench 64SurveyBench –ROASBench 32BulletBench –
11 Claude Sonnet 5.5Anthropic 52 4 of 5 benchmarks CatalogBench 44DeckBench 82SurveyBench –ROASBench 79BulletBench 6
12 Gemini 3.1 Pro PreviewGoogle 49 5 of 5 benchmarks CatalogBench 56DeckBench 18SurveyBench 35ROASBench 54BulletBench 81
13= Gemini 3.8 FlashGoogle 48 4 of 5 benchmarks CatalogBench 38DeckBench 25SurveyBench –ROASBench 68BulletBench 61
13= Gemini 4 ArgonGoogle 48 5 of 5 benchmarks CatalogBench 69DeckBench 36SurveyBench 47ROASBench 64BulletBench 24
15 Kimi K3.1 ProMoonshot AI 45 4 of 5 benchmarks CatalogBench –DeckBench 61SurveyBench 6ROASBench 89BulletBench 24
16= Claude Haiku 5.5Anthropic 42 5 of 5 benchmarks CatalogBench 39DeckBench 50SurveyBench 76ROASBench 7BulletBench 38
16= Particle-1Particle AI 42 4 of 5 benchmarks CatalogBench 6DeckBench 46SurveyBench 65ROASBench 50BulletBench –
18= MiniMax M3.1 FlashMiniMax 39 5 of 5 benchmarks CatalogBench 69DeckBench 71SurveyBench 0ROASBench 43BulletBench 14
20= Kimi K3Moonshot AI 38 4 of 5 benchmarks CatalogBench 37DeckBench 54SurveyBench –ROASBench 57BulletBench 5
20= DeepSeek-V4.1-Flash-1017DeepSeek 38 5 of 5 benchmarks CatalogBench 18DeckBench 43SurveyBench 59ROASBench 46BulletBench 24
22 Mercury 3Inception 37 4 of 5 benchmarks CatalogBench –DeckBench 21SurveyBench 29ROASBench 21BulletBench 76
23 Gemini 3.5 Flash-LiteGoogle 34 5 of 5 benchmarks CatalogBench 26DeckBench 11SurveyBench 24ROASBench 14BulletBench 95
24 Muse Spark 1.3Meta 30 4 of 5 benchmarks CatalogBench 53DeckBench 32SurveyBench –ROASBench 29BulletBench 5
25= Mistral Large 4Mistral AI 27 5 of 5 benchmarks CatalogBench 19DeckBench 7SurveyBench 41ROASBench 36BulletBench 33
25= Qwen3.8-Max (0902)Alibaba 27 3 of 5 benchmarks CatalogBench 23DeckBench 39SurveyBench –ROASBench 18BulletBench –
27 Mistral Medium 3.5Mistral AI 26 5 of 5 benchmarks CatalogBench 33DeckBench 4SurveyBench 18ROASBench 0BulletBench 76
28 Claude Haiku 4.5Anthropic 15 5 of 5 benchmarks CatalogBench 9DeckBench 14SurveyBench 12ROASBench 4BulletBench 35

Not scored yet, with results on fewer than 3 of our benchmarks: DeepSeek-V4-Pro (0423), DeepSeek-V4.1-Flash, GLM-5.3, GLM-5V-Turbo.

How the score works

  1. Percentiles, not raw scores. On each benchmark's headline measure a model scores the share of the other models it beats, from 0 to 100 (ties count half), so a share of products and a chess rating can be averaged. Where a model was run in more than one configuration, such as different reasoning efforts, its best result counts. CatalogBench's two briefs and BulletBench's two clocks are averaged into one percentile each.
  2. The same field of models. Models are compared with the models we tested on CatalogBench or ROASBench, the same field as the overall leaderboard.
  3. Equal weights. The score is the plain average of the percentiles a model has. It needs results on at least 3 of our 5 benchmarks; the coverage column says how many.
  4. What it does not cover. Our benchmarks test business work: product listings, slide decks, survey analysis, ad budgets and fast decisions under time pressure. They do not test coding, agents and tool use, factual knowledge or general chat. A missing benchmark is left out rather than counted as zero, so a model with fewer benchmarks can rank higher than one tested on all of them. Differences of a few points are within noise.

How it differs from the overall leaderboard

The overall leaderboard blends our benchmarks with licensed results from other leaderboards, such as Artificial Analysis and LMArena, weighted towards business work. It covers more ground, but those results are not ours to republish. The Spring Prompt Score uses only results we measured ourselves, so it is open data under CC BY 4.0 and every model can show it as a badge.

Download it as score.csv or score.json from our open data. Each model page has the code to embed its badge. More on the method in our methodology.