Writing fit · public beta
AI model writing-fit standing
A current-domain standing across the writing tasks SpringPrompt has published so far. It is not a general-intelligence leaderboard, and it will broaden as more defensible task composites are published.
GPT 5.5 currently leads the published writing-fit domain.
Current published writing domain
Mean of tie-aware within-task rank percentiles; every published task is required for an overall rank. Raw predicted-fit scores are never averaged across tasks.
Fit data updated 2026-07-16
| Writing standing | Model family | Writing fit | Coverage | Email Writing | Business Writing | Blended price† | Output throughput† |
|---|---|---|---|---|---|---|---|
| #1 | GPT 5.5 OpenAI Operational reference: GPT-5.5 (xhigh) · xhighActive configuration | 96.8mean task percentile | 2/2 | #2fit 62.9 | #2fit 61.8 | $4.35per 1M · 7:2:1 blend | 66.3 tok/safter generation begins |
| #2 | GPT 5.6 Sol OpenAI Operational reference: GPT-5.6 Sol (max) · maxActive configuration | 95.3mean task percentile | 2/2 | #4fit 62.7 | #1fit 63.0 | $4.35per 1M · 7:2:1 blend | 55.9 tok/safter generation begins |
| #3 | Claude Opus 4.8 Anthropic Operational reference: Claude Opus 4.8 (max) · maxActive configuration | 95.0mean task percentile | 2/2 | #1fit 62.9 | #4fit 61.2 | $3.85per 1M · 7:2:1 blend | 54.8 tok/safter generation begins |
| #4 | Claude Fable 5 Anthropic Operational reference: Claude Fable 5 (with fallback) · maxActive configuration | 93.5mean task percentile | 2/2 | #3fit 62.8 | #3fit 61.6 | $7.7per 1M · 7:2:1 blend | 65.0 tok/safter generation begins |
| #5 | GPT 5.4 OpenAI Legacy operational reference: GPT-5.4 (xhigh) · xhighDeprecated · legacy data | 85.5mean task percentile | 2/2 | #6fit 61.9 | #5fit 60.9 | $2.17legacy reference · excluded from filters | 150 tok/slegacy reference · excluded from filters |
| #6 | Claude Opus 4.7 Anthropic Operational reference: Claude Opus 4.7 (max) · maxActive configuration | 85.4mean task percentile | 2/2 | #5fit 62.6 | #6fit 60.7 | $3.85per 1M · 7:2:1 blend | 48.3 tok/safter generation begins |
| #7 | Claude Opus 4.6 Anthropic Legacy operational reference: Claude Opus 4.6 (max) · maxDeprecated · legacy data | 80.6mean task percentile | 2/2 | #7fit 61.5 | #7fit 60.4 | $3.85legacy reference · excluded from filters | 39.0 tok/slegacy reference · excluded from filters |
| #8 | Claude Sonnet 4.6 Anthropic Legacy operational reference: Claude Sonnet 4.6 (max) · maxDeprecated · legacy data | 77.4mean task percentile | 2/2 | #8fit 61.1 | #8fit 59.4 | $2.31legacy reference · excluded from filters | 46.9 tok/slegacy reference · excluded from filters |
| #9 | Claude Sonnet 5 Anthropic Operational reference: Claude Sonnet 5 (max) · maxActive configuration | 74.2mean task percentile | 2/2 | #9fit 59.6 | #9fit 59.2 | $1.54per 1M · 7:2:1 blend | 70.9 tok/safter generation begins |
| #10 | GPT 5.2 OpenAI Legacy operational reference: GPT-5.2 (xhigh) · xhighDeprecated · legacy data | 69.3mean task percentile | 2/2 | #10fit 59.1 | #11fit 58.0 | $1.87legacy reference · excluded from filters | 65.6 tok/slegacy reference · excluded from filters |
Show the other 23 model families ▾
| #11 | Deepseek V4 Pro DeepSeek Operational reference: DeepSeek V4 Pro (max) · maxActive configuration | 66.0mean task percentile | 2/2 | #11fit 58.9 | #12fit 57.8 | $0.18per 1M · 7:2:1 blend | 62.3 tok/safter generation begins |
| #12 | Kimi K2.6 Moonshot AI Operational reference: Kimi K2.6 · reasoningActive configuration | 64.7mean task percentile | 2/2 | #14fit 58.3 | #10fit 58.5 | $0.7per 1M · 7:2:1 blend | 44.4 tok/safter generation begins |
| #13 | Gemini 3 Pro Google | 59.6mean task percentile | 2/2 | #13fit 58.3 | #14fit 57.2 | — | — |
| #14 | GPT 5 OpenAI Legacy operational reference: GPT-5 (high) · highDeprecated · legacy data | 56.6mean task percentile | 2/2 | #16fit 56.8 | #13fit 57.3 | $1.34legacy reference · excluded from filters | 99.0 tok/slegacy reference · excluded from filters |
| #15 | DeepSeek-V4-Flash DeepSeek Operational reference: DeepSeek V4 Flash (max) · maxActive configuration | 53.1mean task percentile | 2/2 | #15fit 57.4 | #16fit 56.9 | $0.06per 1M · 7:2:1 blend | 109 tok/safter generation begins |
| #16 | Grok 4.5 xAI Operational reference: Grok 4.5 (high) · highActive configuration | 50.1mean task percentile | 2/2 | #18fit 56.7 | #15fit 57.0 | $1.35per 1M · 7:2:1 blend | 116 tok/safter generation begins |
| #17 | Gemini 3.1 Pro Google Operational reference: Gemini 3.1 Pro Preview · reasoningActive configuration | 46.7mean task percentile | 2/2 | #17fit 56.8 | #18fit 56.0 | $1.74per 1M · 7:2:1 blend | 116 tok/safter generation begins |
| #18 | Deepseek V3.2 DeepSeek | 45.2mean task percentile | 2/2 | #19fit 55.8 | #17fit 56.1 | — | — |
| #19 | Deepseek V3.1 DeepSeek | 38.8mean task percentile | 2/2 | #21fit 54.8 | #19fit 54.9 | — | — |
| #20 | Kimi K2.5 Moonshot AI Legacy operational reference: Kimi K2.5 · reasoningDeprecated · legacy data | 38.6mean task percentile | 2/2 | #20fit 55.4 | #20fit 54.9 | $0.49legacy reference · excluded from filters | 49.8 tok/slegacy reference · excluded from filters |
| #21 | Qwen3.5-397B-A17B Alibaba Operational reference: Qwen3.5 397B A17B · reasoningActive configuration | 33.9mean task percentile | 2/2 | #22fit 53.8 | #21fit 53.1 | $0.9per 1M · 7:2:1 blend | 59.6 tok/safter generation begins |
| #22 | GPT 4.1 OpenAI Legacy operational reference: GPT-4.1 · non-reasoningDeprecated · legacy data | 30.6mean task percentile | 2/2 | #23fit 52.4 | #22fit 52.7 | $1.55legacy reference · excluded from filters | 96.0 tok/slegacy reference · excluded from filters |
| #23 | DeepSeek-R1-0528 DeepSeek | 27.4mean task percentile | 2/2 | #24fit 51.8 | #23fit 51.8 | — | — |
| #24 | DeepSeek-R1 DeepSeek | 24.2mean task percentile | 2/2 | #25fit 51.1 | #24fit 51.5 | — | — |
| #25 | DeepSeek-V3-0324 DeepSeek | 19.4mean task percentile | 2/2 | #27fit 49.3 | #25fit 50.4 | — | — |
| #26 | gemma-3-12b-it Google | 17.6mean task percentile | 2/2 | #26fit 49.3 | #27fit 49.0 | — | — |
| #27 | gemma-3-27b-it Google | 16.1mean task percentile | 2/2 | #28fit 48.9 | #26fit 49.6 | — | — |
| #28 | qwq-32b Alibaba | 11.2mean task percentile | 2/2 | #29fit 47.1 | #28fit 47.4 | — | — |
| #29 | gemma-3-4b-it Google | 8.0mean task percentile | 2/2 | #30fit 45.2 | #29fit 46.4 | — | — |
| #30 | gpt-oss-120b OpenAI Operational reference: gpt-oss-120b (high) · highActive configuration | 4.8mean task percentile | 2/2 | #31fit 44.6 | #30fit 45.5 | $0.2per 1M · 7:2:1 blend | 255 tok/safter generation begins |
| #31 | gpt-oss-20b OpenAI Operational reference: gpt-oss-20b (high) · highActive configuration | 1.6mean task percentile | 2/2 | #32fit 41.8 | #31fit 43.0 | $0.07per 1M · 7:2:1 blend | 205 tok/safter generation begins |
| — | GPT 5.1 OpenAI · provisional coverage Legacy operational reference: GPT-5.1 (high) · highDeprecated · legacy data | 65.6mean task percentile | 1/2not overall-ranked | #12fit 58.4 | — | $1.34legacy reference · excluded from filters | 106 tok/slegacy reference · excluded from filters |
| — | gemma-2-9b-it Google · provisional coverage | 0.0mean task percentile | 1/2not overall-ranked | #33fit 41.5 | — | — | — |
Full fit coverage means appearing in all 2 currently published task cohorts. A thinner row remains visible but receives no overall fit position. †Operational coverage is 14 active, 8 legacy, and 11 missing out of 33 model families. 14 active references have blended price data and 14 have output-throughput data. Legacy values remain visible for context but are excluded from cheapest and fastest filters. Neither metric affects writing fit.
Why percentiles?
Each task has its own evidence mix, so its raw predicted-fit numbers are local to that page. Tie-aware rank percentiles put positions onto a common scale without pretending the raw scores are interchangeable.
What cost and speed mean
Blended price uses Artificial Analysis's 7:2:1 cache-hit, uncached-input, and output-token weights per million tokens. Output throughput is the source-listed median after generation begins; it excludes time to first token and is not end-to-end response latency. Only active reference configurations are eligible for cheapest or fastest positions; deprecated references are labelled as legacy data. These metrics help with deployment choices; they are not quality points.
Frequently asked
Is this a general AI intelligence leaderboard?
No. It is the current SpringPrompt writing-fit standing across only the published writing-domain task composites shown on this page.
How is the writing-fit standing calculated?
Each task is converted to a tie-aware within-task rank percentile. Those percentiles are averaged only for models covered by every published task; raw predicted-fit scores are never averaged across tasks.
Do price or output speed affect the standing?
No. Source-listed blended price and median output throughput are shown as separate operational estimates and never change a model's writing-fit position. Only active reference configurations are eligible for cheapest or fastest positions; deprecated references are labelled as legacy data.