WritingBench

We show this table for reference; we do not rank on it.

WritingBench evaluates generative writing across six domains. This dated benchmark-owner table contains 54 source model rows, with overall and domain scores on the native 0-100 scale.

Evaluation results published by WritingBench / X-PLUG project. Published overall and six domain aggregates on a 0-100 point scale. The 1,000-query count applies to each overall row; domain counts are not reported.

WritingBench: A Comprehensive Benchmark for Generative Writing, WritingBench / X-PLUG project; official aggregate workbook. Apache 2.0 license. Source and transformation notice. Published aggregate workbook only. No task texts, evaluation prompts, model responses, or model weights are included.

Writing score on WritingBench — November 27, 2025 source update

We mirror the published writing score view for WritingBench. gpt-5-2025-08-07 leads the public snapshot at 83.87, followed by qwen3-235b-a22b-thinking-2507 (82.34) and kimi-k2-instruct (81.26). We do not use these results to rank models overall.

gpt-5-2025-08-07

OpenAI

83.87

2Open weights

qwen3-235b-a22b-thinking-2507

Alibaba

82.34

Context 128K

kimi-k2-instruct

Moonshot

81.26

54 source model rowsInstruction FollowingDated writing cohortDisplay onlyUpdated November 27, 2025 source update

WritingBench, November 2025 cohort (54 source model rows)
Source model	Overall	Academic & Engineering	Finance & Business	Politics & Law	Literature & Arts	Education	Advertising & Marketing	Parameters (B)	Open / closed
gpt-5-2025-08-07OpenAI	83.87	84.46	83.78	82.21	85.26	85.64	82.34	Not reported	Unknown
qwen3-235b-a22b-thinking-2507Alibaba	82.34	83.43	83.44	79.67	82.68	84.52	80.94	Not reported	Open
kimi-k2-instructMoonshot	81.26	82.23	81.90	80.32	81.73	80.97	80.02	Not reported	Unknown
claude-sonnet-4-5-20250929Anthropic	80.71	80.77	80.54	79.59	81.26	82.32	80.45	Not reported	Unknown
o3-2025-04-16OpenAI	80.46	80.83	81.15	78.31	80.47	81.59	81.23	Not reported	Unknown
qwen3-235b-a22b-instruct-2507Alibaba	80.26	80.80	81.14	79.14	80.47	80.97	78.97	Not reported	Open
qwen3-maxAlibaba	80.26	80.46	80.10	78.70	81.50	81.75	79.70	Not reported	Unknown
claude-opus-4-5-20251101Anthropic	79.78	79.31	80.00	78.05	81.41	81.30	79.11	Not reported	Unknown
doubao-seed-1-6-thinking-250715ByteDance	79.29	80.79	79.74	78.05	78.37	81.48	77.95	Not reported	Unknown
gemini-2.5-proGoogle	79.26	79.59	78.24	78.69	80.92	81.30	77.27	Not reported	Unknown
nanbeige4-3b-thinking-2511Boss	79.03	80.77	82.37	77.41	76.05	80.81	76.55	Not reported	Unknown
deepseek-r1-0528Deepseek	78.92	78.79	78.38	77.27	80.91	80.41	78.44	Not reported	Unknown
doubao-seed-1-6-thinking-250615ByteDance	78.76	80.07	79.99	76.98	77.06	81.73	77.72	Not reported	Unknown
gemini-3-pro-previewGoogle	78.50	77.13	77.26	77.55	81.11	80.58	78.27	Not reported	Unknown
doubao-seed-1-6-250615ByteDance	78.33	80.18	78.80	72.90	79.91	81.06	79.02	Not reported	Unknown
claude-haiku-4-5-20251001Anthropic	77.09	77.54	78.61	75.33	76.35	79.62	75.59	Not reported	Unknown
claude-opus-4-1-20250805Anthropic	76.36	76.89	75.56	74.19	78.05	77.33	77.16	Not reported	Unknown
claude-opus-4-20250514-thinkingAnthropic	75.12	75.60	74.57	73.09	77.20	75.78	74.98	Not reported	Unknown
qwen3-235b-a22b-thinkingAlibaba	74.73	75.45	74.84	73.30	74.67	76.41	74.50	Not reported	Unknown
grok-4-0709xAI	74.65	75.44	73.99	73.08	74.66	77.48	74.73	Not reported	Unknown
claude-opus-4-20250514Anthropic	74.45	73.90	73.96	73.05	76.31	75.23	74.81	Not reported	Unknown
claude-sonnet-4-20250514Anthropic	74.10	75.35	74.13	71.55	74.66	74.99	74.84	Not reported	Unknown
claude-sonnet-4-20250514-thinkingAnthropic	74.07	75.08	74.52	70.71	75.26	76.04	73.88	Not reported	Unknown
claude-3-7-sonnet-20250219-thinkingAnthropic	73.40	73.11	73.12	72.43	73.93	74.43	74.11	Not reported	Unknown
chatgpt-4o-latestOpenAI	73.22	72.79	73.08	72.16	73.58	73.71	74.77	Not reported	Unknown
o4-mini-2025-04-16OpenAI	72.90	76.04	73.19	71.26	69.87	75.91	72.66	Not reported	Unknown
claude-3-7-sonnet-20250219Anthropic	72.45	72.46	72.78	70.80	71.69	74.74	73.56	Not reported	Unknown
qwen3-235b-a22b-nothinkingAlibaba	72.44	72.40	72.78	71.36	72.67	74.25	71.70	Not reported	Unknown
qwq-32bAlibaba	72.40	72.92	72.04	71.71	70.45	75.33	73.66	Not reported	Unknown
deepseek-r1DeepSeek	72.13	71.95	70.35	71.52	71.93	74.54	74.45	Not reported	Unknown
grok-3xAI	71.72	71.40	71.73	72.19	71.14	74.68	69.63	Not reported	Unknown
gpt-4.1-2025-04-14OpenAI	70.90	71.52	70.26	70.21	70.45	72.20	71.77	Not reported	Closed
deepseek-v3-250324DeepSeek	69.36	68.97	68.13	68.27	69.78	71.21	71.41	Not reported	Unknown
o3-miniOpenAI	68.02	69.52	68.61	66.92	66.54	71.48	65.95	Not reported	Unknown
gpt-4.5-previewOpenAI	67.61	68.22	66.17	66.83	66.40	70.56	69.59	Not reported	Unknown
gemini-2.0-flashGoogle	66.80	67.51	65.78	65.84	66.24	69.53	67.50	Not reported	Unknown
qwen-maxAlibaba	66.78	66.37	65.52	65.22	67.36	69.77	68.44	Not reported	Unknown
o1-preview-2024-09-12OpenAI	63.44	63.56	62.30	62.37	63.18	65.53	65.39	Not reported	Unknown
deepseek-v3DeepSeek	59.23	59.64	57.86	58.32	57.85	62.34	61.67	Not reported	Unknown
gemini-1.5-proGoogle	58.80	58.83	55.69	57.22	60.33	61.78	61.61	Not reported	Unknown
qwen-2.5-72b-instructAlibaba	57.78	59.37	56.93	57.27	55.33	62.25	57.53	Not reported	Unknown
gpt-4o-miniOpenAI	56.40	57.14	54.69	54.34	56.49	59.05	59.03	Not reported	Unknown
mistral-large-instruct-2411Mistral AI	54.44	55.54	54.35	53.73	51.67	57.30	55.77	Not reported	Unknown
claude-3-5-sonnet-20241022Anthropic	53.76	52.77	52.44	51.69	53.10	55.78	59.69	Not reported	Unknown
llama-4-maverick-17b-128eMeta	52.55	51.83	51.58	51.35	51.25	55.42	56.36	Not reported	Unknown
llama-4-scout-17b-16e-instructMeta	52.02	52.57	52.05	51.47	49.19	54.76	53.78	Not reported	Unknown
qwen-2.5-7b-instructAlibaba	49.46	51.86	50.43	49.21	43.79	53.59	49.64	Not reported	Unknown
claude-3-5-haiku-20241022Anthropic	49.06	49.70	48.72	47.33	46.25	50.20	54.52	Not reported	Unknown
llama-3.3-70b-instructMeta	48.04	47.58	48.52	45.73	46.04	50.86	51.84	Not reported	Unknown
qwen3-0.6b-thinkingAlibaba	45.08	50.13	47.28	45.65	37.98	47.46	42.06	Not reported	Unknown
qwen3-0.6b-nothinkingAlibaba	41.47	46.37	43.52	41.85	34.61	42.81	39.72	Not reported	Unknown
llama-3.1-8b-instructMeta	40.18	41.58	41.49	37.02	36.52	42.92	44.02	Not reported	Unknown
llama-3.2-3b-instructMeta	38.14	40.26	39.89	34.78	34.27	41.01	40.84	Not reported	Unknown
llama-3.2-1b-instructMeta	28.74	32.56	30.03	25.44	25.18	30.11	30.73	Not reported	Unknown

All seven columns use native 0-100 score points, displayed to two decimals. The 1,000-query total is not each domain's denominator. The source does not report generation settings or individual run dates. This table does not move overall or category rankings.

How we show WritingBench

We extracted all 54 model rows from the official aggregate workbook at the pinned November 27, 2025 source update. This historical table names Claude Sonnet 4.5 as the judge. Its update date is separate from individual evaluation dates, which are not reported.

Overall and domain values retain their native 0-100 scale. They are judge scores, not accuracy percentages. The source already multiplied its original 10-point assessments by ten; we do not repeat that scaling or recompute the published overall score from domain averages.

Three profile references match reviewed model versions and modes. The other 51 source rows keep their own identities. Generation settings, thinking budgets, endpoints, and weight revisions remain unknown, including for the three linked models. These results are excluded from overall and category rankings.

Snapshot

1,000 queriesSix domains0-100 pointsNovember 2025 cohortDisplay only

Official aggregate workbook Source license declaration Apache 2.0 license copy Source and transformation notice

The published WritingBench snapshot places gpt-5-2025-08-07 first at 83.87. The third row is 2.61 score units behind. The broader top-10 range is 4.61 score units, so many of the published results sit in a relatively narrow band.

54 source model rows are shown for WritingBench. The benchmark falls in the Instruction Following category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. WritingBench is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.

About WritingBench

Year

2025

Tasks

1,000 queries

Format

Judge score, 0-100 points

The source update is dated November 27, 2025 and names Claude Sonnet 4.5 as the judge. Individual run dates, generation parameters, thinking budgets, and domain counts are not reported in the workbook. This historical cohort stays display only and does not enter overall or category rankings.

WritingBench: A Comprehensive Benchmark for Generative Writing Public benchmark source

Freshness and provenance

Version

November 27, 2025 aggregate workbook

Refresh cadence

Pinned source artifact

Staleness state

Dated writing cohort

Question availability

Public upstream; no task contents mirrored

Dated writing cohortDisplay only

BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.

Questions

What does WritingBench measure?

WritingBench evaluates generative writing across six domains. This dated benchmark-owner table contains 54 source model rows, with overall and domain scores on the native 0-100 scale.

Which model leads the published WritingBench snapshot?

gpt-5-2025-08-07 currently leads the published WritingBench snapshot with 83.87 writing score. BenchLM shows this benchmark for display only and does not use it in overall rankings.

How many models are evaluated on WritingBench?

The November 27, 2025 source update snapshot contains 54 source model rows.

Are WritingBench scores accuracy percentages?

The workbook reports judge scores on a 0-100 scale. We keep its native values and show overall and domain scores separately. The source already scaled its original 10-point assessments before publishing this table. Applying that scaling again would inflate the values tenfold.

Do the profile links identify exact evaluation settings?

Three profile references match the reviewed model version and mode. The remaining source rows stay separate. Even those three links do not establish an endpoint, weight revision, quantization, temperature, output cap, or thinking budget. The source update date also does not establish an individual run date.

Last updated: November 27, 2025 source update · mirrored from the public benchmark leaderboard

Know when it’s worth switching models

The model to choose, the cheaper alternative, and the release we would wait on.

Read a sample issue

Join 5,600+ readers.

One email each week. Unsubscribe anytime.