Skip to content
assaynotes
EN日本語中文
Evaluation sources

Creative Writing v3

EQ-Bench · Short-form creation and expression

Official evaluation
In NewsRanked
Evidence budget3.6%
Upstream data as of10/29 02:36
Last synced11/16 02:05

What it measures, and how

The official CSV is embedded in the data script and uses Elo; Sonnet 4.6 judges wins and losses, and style, repetition rate and rubric are not counted twice.

How this evidence is used

Officially scored; the two writing leaderboards together account for 6%, with no added weight by question or model count.

Results

Under fixed rules, each public model uses one representative configuration. Anonymous test variants are not shown, original leaderboard ranks are kept, and each item lists at most 30.

Rank thereModel thereRaw scoreRepresentative configuration
1gpt-6-astraOpenAI2,176.9Source default configuration
2kimi-k3.1-promoonshot2,172.6Source default configuration
3claude-fable-5-1Anthropic2,155.3Source default configuration
4claude-opus-5Anthropic2,135Source default configuration
5gpt-6-solOpenAI2,126Source default configuration
6kimi-k3Moonshot AI2,083.3Source default configuration
7GLM-5.3Z.ai2,076.4Source default configuration
8claude-opus-5-5Anthropic2,050.9Source default configuration
9grok-4.7xAI2,012.2Source default configuration
10muse-wondermeta2,007Source default configuration
11ox-alpha—1,973.7Source default configuration
12gpt-5.6-solOpenAI1,973.5Source default configuration
13claude-fable-5Anthropic1,943.6Source default configuration
14aion-3.5aion-labs1,930.6Source default configuration
15muse-spark-1.1Meta1,928.2Source default configuration
16claude-opus-4-7Anthropic1,914.9Source default configuration
17Altworld/Hemmingway-1—1,907.4Source default configuration
18muse-spark-1.3Meta1,906.9Source default configuration
19gpt-5.6-terraOpenAI1,856.2Source default configuration
20gpt-5.5OpenAI1,844.5Source default configuration
21Qwen/Qwen3.8-2.4T-A95BAlibaba1,843.8Source default configuration
22muse-spark-1.2Meta1,842.2Source default configuration
23gpt-5.4OpenAI1,840.8Source default configuration
24claude-opus-4-8Anthropic1,840.1Source default configuration
25gpt-5.6-lunaOpenAI1,831.5Source default configuration
26claude-sonnet-4-6Anthropic1,811Source default configuration
27claude-opus-4-6Anthropic1,810.3Source default configuration
28meta-models/Muse-Glimmer-30BMeta1,804.3Source default configuration
29claude-sonnet-5Anthropic1,797.9Source default configuration
30space-bunny-alphastealth1,788.9Source default configuration
Limits and data attribution

Mainly English writing, relying on a model judge; the same judge and related tasks share a budget and do not represent Chinese prose quality.

Data licence: MIT · EQ-bench-site README metadata statement

Scores published by EQ-Bench. Raw scores and the News consensus score use different scales and cannot be added directly.