> ## Documentation Index
> Fetch the complete documentation index at: https://usefulai.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Best LLMs for Writing in 2026

> Compare the best LLMs for writing in 2026 by quality, price, and access, with picks for creative work, editing, research-backed prose, and local use.

Updated July 12, 2026

The best LLM for writing isn't the one topping general leaderboards - writing quality and reasoning quality often diverge. We ranked 15 models by a human-preference creative-writing benchmark, then added judgment on price, access, and where each one actually earns its slot.

## Best LLMs for Writing

| # | Model | Best for | Score About score (UsefulAI's 0-100 writing score is normalized from Arena Text Creative Writing Elo. Higher reflects stronger human preference for creative prose.) | Price About price (Estimated blended API price per 1M tokens using a writing-heavy input-to-output ratio. Tiers and route differences can change the real cost.) | License About license (Proprietary means no public model weights. Open weight means weights are available, though exact licenses and commercial-use terms vary.) |
| -: | - | - | -: | -: | - |
| 1 | [![](https://usefulai.com/images/icons/48/anthropic.com.png)Claude Fable 5](https://www.anthropic.com/news/claude-fable-5-mythos-5) | Top-tier creative prose | 100% | \$40.00 / 1M | Proprietary |
| 2 | [![](https://usefulai.com/images/icons/48/anthropic.com.png)Claude Opus 4.6](https://www.anthropic.com/news/claude-opus-4-6) | High-end prose value | 98% | \$20.00 / 1M | Proprietary |
| 3 | [![](https://usefulai.com/images/icons/48/google.com.png)Gemini 3.1 Pro](https://ai.google.dev/gemini-api/docs/models/gemini-3.1-pro-preview) | Long-form research writing | 79% | \$9.50 / 1M | Proprietary |
| 4 | [![](https://usefulai.com/images/icons/48/google.com.png)Gemini 3.5 Flash](https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash) | High-volume drafting and editing | 72% | \$7.13 / 1M | Proprietary |
| 5 | [![](https://usefulai.com/images/icons/48/anthropic.com.png)Claude Opus 4.8](https://www.anthropic.com/news/claude-opus-4-8) | Current flagship Claude default | 64% | \$20.00 / 1M | Proprietary |
| 6 | [![](https://usefulai.com/images/icons/48/meta.ai.png)Muse Spark](https://ai.meta.com/blog/introducing-muse-spark-msl/) | Writing inside Meta AI | 64% | n/a | Proprietary |
| 7 | [![](https://usefulai.com/images/icons/48/x.ai.png)Grok 4.20](https://docs.x.ai/developers/models) | Distinctive voice and style | 60% | \$2.19 / 1M | Proprietary |
| 8 | [![](https://usefulai.com/images/icons/48/openai.com.png)GPT-5.5](https://openai.com/index/gpt-5-5-instant) | Reliable general-purpose writing | 54% | \$23.75 / 1M | Proprietary |
| 9 | [![](https://usefulai.com/images/icons/48/qwen.ai.png)Qwen3.7 Max](https://qwen.ai/blog?id=qwen3.7) | Long-context multilingual writing | 44% | \$3.13 / 1M | Proprietary |
| 10 | [![](https://usefulai.com/images/icons/48/z.ai.png)GLM-5.2](https://docs.z.ai/guides/llm/glm-5.2) | Current open-weight writing | 44% | \$3.65 / 1M | Open weight |
| 11 | [![](https://usefulai.com/images/icons/48/deepseek.com.png)DeepSeek V4 Pro](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro) | Low-cost open-weight writing | 42% | \$0.76 / 1M | Open weight |
| 12 | [![](https://usefulai.com/images/icons/48/mimo.xiaomi.com.png)MiMo-V2.5-Pro](https://mimo.xiaomi.com/mimo-v2-5-pro/) | Cheap open-weight drafting | 35% | \$0.76 / 1M | Open weight |
| 13 | [![](https://usefulai.com/images/icons/48/anthropic.com.png)Claude Sonnet 5](https://www.anthropic.com/news/claude-sonnet-5) | Everyday Claude writing value | 32% | \$8.00 / 1M | Proprietary |
| 14 | [![](https://usefulai.com/images/icons/48/kimi.com.png)Kimi K2.6](https://platform.kimi.ai/docs/guide/kimi-k2-6-quickstart) | Open-weight prose and critique | 31% | \$3.24 / 1M | Open weight |
| 15 | [![](https://usefulai.com/images/icons/48/google.com.png)Gemma 4 31B](https://huggingface.co/google/gemma-4-31B-it) | Laptop-friendly local writing | 20% | \$0.34 / 1M | Open weight |

Show more
Show less

***

![](https://mintcdn.com/usefulai/tO2qspLJNjFc61Zv/images/icons/144/anthropic.com.png?fit=max\&auto=format\&n=tO2qspLJNjFc61Zv\&q=85\&s=2077996fd7746bfe8ee85acdd8018de3)

## [Claude Fable 5](https://www.anthropic.com/news/claude-fable-5-mythos-5)

Anthropic

Top-tier creative prose

[Visit Anthropic](https://www.anthropic.com/news/claude-fable-5-mythos-5)

The strongest pure prose model in the current benchmark, and it shows most on fiction, voice, and nuance, though you pay a real premium for it.

Score **100%**
Price **\$40.00 / 1M**
License Proprietary
Context **1M**

Strengths

<ul>
  <li>It sits at the top of our writing benchmark for a reason: it holds a consistent voice across long pieces, handles subtext and rhythm, and rarely flattens into generic AI cadence.</li>
  <li>If you want the highest ceiling for fiction, essays, or brand voice, this is the pick.</li>
</ul>

Tradeoffs

<ul>
  <li>It's the most expensive model here by a wide margin, and creative-writing strength doesn't guarantee factual accuracy or clean SEO structure.</li>
  <li>For research-heavy or high-volume drafting, Gemini 3.1 Pro or Claude Sonnet 5 give you most of the quality for far less.</li>
</ul>

How to access

<ul>
  <li>**App** — Available in [Claude](https://claude.ai/).</li>
  <li>**API** — Accessible via [Anthropic API](https://docs.anthropic.com/en/api/overview).</li>
</ul>

![](https://mintcdn.com/usefulai/tO2qspLJNjFc61Zv/images/icons/144/anthropic.com.png?fit=max\&auto=format\&n=tO2qspLJNjFc61Zv\&q=85\&s=2077996fd7746bfe8ee85acdd8018de3)

## [Claude Opus 4.6](https://www.anthropic.com/news/claude-opus-4-6)

Anthropic

High-end prose value

[Visit Anthropic](https://www.anthropic.com/news/claude-opus-4-6)

A previous-generation Opus that still writes near the very top of this list for meaningfully less than Fable 5.

Score **98%**
Price **\$20.00 / 1M**
License Proprietary
Context **1M**

Strengths

<ul>
  <li>It nearly matches Fable 5 on prose quality for far less, which makes it our value pick for serious writing. It's strong at long-form structure, argument, and holding tone, and it writes better than the newer Opus 4.8.</li>
  <li>Reach for it when you want near-frontier output without the top price.</li>
</ul>

Tradeoffs

<ul>
  <li>Being a prior release, it may be retired or repriced before newer Claudes, so check availability if you're building on it.</li>
  <li>Fable 5 still has a higher ceiling for the hardest creative work, and Opus 4.8 is the stronger all-round reasoning model.</li>
</ul>

How to access

<ul>
  <li>**App** — Available in [Claude](https://claude.ai/).</li>
  <li>**API** — Accessible via [Anthropic API](https://docs.anthropic.com/en/api/overview).</li>
</ul>

![](https://mintcdn.com/usefulai/Ez-pJDkPpztLE7Cr/images/icons/144/google.com.png?fit=max\&auto=format\&n=Ez-pJDkPpztLE7Cr\&q=85\&s=11e0c725f841c45443116a184b63beac)

## [Gemini 3.1 Pro](https://ai.google.dev/gemini-api/docs/models/gemini-3.1-pro-preview)

Google

Long-form research writing

[Visit Google](https://ai.google.dev/gemini-api/docs/models/gemini-3.1-pro-preview)

Google's strongest current writing model, best when your draft leans on long source documents and research rather than pure style.

Score **79%**
Price **\$9.50 / 1M**
License Proprietary
Context **1M**

Strengths

<ul>
  <li>The best pick here for research-heavy and long-form work, with a huge context that lets you draft from many sources at once. It stays organized across long outputs and handles structured, factual writing better than most of the higher-scoring creative models.</li>
  <li>A strong default for reports and documentation.</li>
</ul>

Tradeoffs

<ul>
  <li>It's a notch below the top Claude models on voice and creative nuance, so it's not our first choice for fiction or distinctive brand writing.</li>
  <li>It's also a preview release on Google's routes, so pricing and availability can shift - confirm both before you commit.</li>
</ul>

How to access

<ul>
  <li>**App** — Available in [Gemini](https://gemini.google.com/).</li>
  <li>**API** — Accessible via [Gemini API](https://ai.google.dev/gemini-api/docs/models/gemini-3.1-pro-preview) and [Gemini Enterprise Agent Platform](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-1-pro).</li>
</ul>

![](https://mintcdn.com/usefulai/Ez-pJDkPpztLE7Cr/images/icons/144/google.com.png?fit=max\&auto=format\&n=Ez-pJDkPpztLE7Cr\&q=85\&s=11e0c725f841c45443116a184b63beac)

## [Gemini 3.5 Flash](https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash)

Google

High-volume drafting and editing

[Visit Google](https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash)

A fast, cheaper Gemini that punches above its tier for writing, ideal when you're generating or editing at volume.

Score **72%**
Price **\$7.13 / 1M**
License Proprietary
Context **1M**

Strengths

<ul>
  <li>Unusually strong prose for a fast, low-cost model, with the same large context as the Pro tier.</li>
  <li>It's built for speed and throughput, making it the pick when you're drafting, rewriting, or editing in bulk and want quality that holds up without slowing you down.</li>
</ul>

Tradeoffs

<ul>
  <li>It doesn't have the top prose ceiling, so for your most important creative or high-stakes pieces, Fable 5, Opus 4.6, or Gemini 3.1 Pro do better.</li>
  <li>Use Flash where volume, cost, and speed matter more than peak polish.</li>
</ul>

How to access

<ul>
  <li>**App** — Available in [Gemini](https://gemini.google.com/).</li>
  <li>**API** — Accessible via [Gemini API](https://ai.google.dev/gemini-api/docs).</li>
</ul>

![](https://mintcdn.com/usefulai/tO2qspLJNjFc61Zv/images/icons/144/anthropic.com.png?fit=max\&auto=format\&n=tO2qspLJNjFc61Zv\&q=85\&s=2077996fd7746bfe8ee85acdd8018de3)

## [Claude Opus 4.8](https://www.anthropic.com/news/claude-opus-4-8)

Anthropic

Current flagship Claude default

[Visit Anthropic](https://www.anthropic.com/news/claude-opus-4-8)

The newest flagship Opus and a superb all-round model, though for pure creative writing the older Opus 4.6 actually scores higher.

Score **64%**
Price **\$20.00 / 1M**
License Proprietary
Context **1M**

Strengths

<ul>
  <li>The most capable current Claude for reasoning, instruction-following, and mixed work that blends writing with analysis or code.</li>
  <li>When your writing sits inside broader tasks - briefs, technical docs, judgment-heavy editing - it's a dependable default that keeps quality high across the whole job.</li>
</ul>

Tradeoffs

<ul>
  <li>For pure creative prose it's a real step down from Opus 4.6, an unusual case where the older Opus is the better writer at the same price.</li>
  <li>If style and voice are your priority, choose Opus 4.6 or Fable 5 instead.</li>
</ul>

How to access

<ul>
  <li>**App** — Available in [Claude](https://claude.ai/).</li>
  <li>**API** — Accessible via [Anthropic API](https://docs.anthropic.com/en/api/overview) and [Amazon Bedrock](https://aws.amazon.com/blogs/machine-learning/claude-opus-4-8-is-now-available-on-aws/).</li>
</ul>

![](https://mintcdn.com/usefulai/52KaILwddzz5TNZ_/images/icons/144/meta.ai.png?fit=max\&auto=format\&n=52KaILwddzz5TNZ_\&q=85\&s=67eba125dab2438bdf2c8a33074498c7)

## [Muse Spark](https://ai.meta.com/blog/introducing-muse-spark-msl/)

Meta

Writing inside Meta AI

[Visit Meta](https://ai.meta.com/blog/introducing-muse-spark-msl/)

Meta's surprise strong writer, but you can only use it inside the Meta AI app - there's no API and no pricing to compare.

Score **64%**
Price **n/a**
License Proprietary
Context **n/a**

Strengths

<ul>
  <li>Genuinely strong prose that scores with the mid-pack frontier models, wrapped in a free, mainstream consumer app.</li>
  <li>If you write casually and just want good output without setting up API access or paying per token, it's an easy and capable option.</li>
</ul>

Tradeoffs

<ul>
  <li>There's no public API or published pricing, so you can't build on it or budget for it, and there's no local route.</li>
  <li>For any programmatic or team workflow, a model with real API access - the Claude, Gemini, or GPT options - is the practical choice.</li>
</ul>

How to access

<ul>
  <li>**App** — Available in [Meta AI](https://www.meta.ai/).</li>
</ul>

![](https://mintcdn.com/usefulai/sV7VJe4pqO2Le0pu/images/icons/144/x.ai.png?fit=max\&auto=format\&n=sV7VJe4pqO2Le0pu\&q=85\&s=421f0adc6c1e753bf2ad0cd1661abbc7)

## [Grok 4.20](https://docs.x.ai/developers/models)

xAI

Distinctive voice and style

[Visit xAI](https://docs.x.ai/developers/models)

The Grok to use for writing with edge and personality, and notably it beats the newer Grok 4.3 on prose.

Score **60%**
Price **\$2.19 / 1M**
License Proprietary
Context **1M**

Strengths

<ul>
  <li>It writes with more attitude and less hedging than most models here, which makes it fun for opinionated, casual, or voice-driven content.</li>
  <li>It's also cheap for its quality, so it's a sensible pick when you want personality and volume without a big bill.</li>
</ul>

Tradeoffs

<ul>
  <li>That distinctive voice can tip into glib or off-tone for formal and professional writing, where a Claude or Gemini model is safer.</li>
  <li>Worth knowing that the newer Grok 4.3 writes worse here, so don't assume the higher version number is the better writer.</li>
</ul>

How to access

<ul>
  <li>**App** — Available in [Grok](https://grok.com/).</li>
  <li>**API** — Accessible via [xAI API](https://docs.x.ai/developers/models/grok-4.20).</li>
</ul>

![](https://mintcdn.com/usefulai/Te6KzZ86-OxPuEC2/images/icons/144/openai.com.png?fit=max\&auto=format\&n=Te6KzZ86-OxPuEC2\&q=85\&s=745b8837f7535bc53cd70fc2f7024d58)

## [GPT-5.5](https://openai.com/index/gpt-5-5-instant)

OpenAI

Reliable general-purpose writing

[Visit OpenAI](https://openai.com/index/gpt-5-5-instant)

OpenAI's best writer in our set and a dependable generalist, though it trails the top Claude and Gemini models on prose.

Score **54%**
Price **\$23.75 / 1M**
License Proprietary
Context **1M**

Strengths

<ul>
  <li>A well-rounded, familiar writing model that handles most everyday tasks - drafts, emails, summaries, rewrites - with steady quality and strong instruction-following.</li>
  <li>If you want one broadly capable model for mixed writing work, it's an easy and low-risk default.</li>
</ul>

Tradeoffs

<ul>
  <li>It sits mid-pack for creative prose, so for fiction, voice, or your highest-stakes pieces the top Claude and Gemini models clearly do better.</li>
  <li>Note that our score is for the Instant variant, so higher-effort GPT-5.5 modes may read differently.</li>
</ul>

How to access

<ul>
  <li>**App** — Available in [ChatGPT](https://chatgpt.com/).</li>
  <li>**API** — Accessible via [OpenAI API](https://developers.openai.com/api/docs/models).</li>
</ul>

![](https://mintcdn.com/usefulai/Te6KzZ86-OxPuEC2/images/icons/144/qwen.ai.png?fit=max\&auto=format\&n=Te6KzZ86-OxPuEC2\&q=85\&s=77ec239e207f3d7865895d11af129b39)

## [Qwen3.7 Max](https://qwen.ai/blog?id=qwen3.7)

Alibaba

Long-context multilingual writing

[Visit Alibaba](https://qwen.ai/blog?id=qwen3.7)

Alibaba's flagship writer, most interesting for long-context and multilingual drafting rather than top-tier English prose.

Score **44%**
Price **\$3.13 / 1M**
License Proprietary
Context **1M**

Strengths

<ul>
  <li>A capable flagship with a large context window and solid multilingual range, useful if you write across languages or from long documents.</li>
  <li>It's reasonably priced for a proprietary frontier model, which makes it a practical choice when breadth matters more than peak English prose.</li>
</ul>

Tradeoffs

<ul>
  <li>For English creative writing it sits mid-pack, well behind the Claude and Gemini leaders.</li>
  <li>It's also a preview release with regional access and API caveats, so confirm availability in your region before you rely on it.</li>
</ul>

How to access

<ul>
  <li>**App** — Available in [Qwen Chat](https://qwen.ai/).</li>
  <li>**API** — Accessible via [Alibaba Cloud Model Studio](https://www.alibabacloud.com/help/en/model-studio/models).</li>
</ul>

![](https://mintcdn.com/usefulai/sV7VJe4pqO2Le0pu/images/icons/144/z.ai.png?fit=max\&auto=format\&n=sV7VJe4pqO2Le0pu\&q=85\&s=d720f43ee25f6663eb6cb5ad076ac3ad)

## [GLM-5.2](https://docs.z.ai/guides/llm/glm-5.2)

Z.ai

Current open-weight writing

[Visit Z.ai](https://docs.z.ai/guides/llm/glm-5.2)

The current GLM and a solid open-weight writer, though the older GLM-5.1 actually scores higher for prose.

Score **44%**
Price **\$3.65 / 1M**
License Open weight
Context **1M**

Strengths

<ul>
  <li>A strong open-weight option with a large context and weights you can route through whichever host fits your cost or compliance needs.</li>
  <li>Good all-round writing quality for the price, and it's the most current, best-supported model in the GLM line.</li>
</ul>

Tradeoffs

<ul>
  <li>The older GLM-5.1 writes better on this benchmark, so pick 5.2 for freshness and support, not for top score.</li>
  <li>Despite open weights, it needs real self-hosting infrastructure rather than a laptop, so most people will use a hosted API anyway.</li>
</ul>

How to access

<ul>
  <li>**App** — Available in [Z.ai](https://chat.z.ai/).</li>
  <li>**API** — Accessible via [Z.ai API](https://docs.z.ai/).</li>
  <li>**Run locally** — Open weights are available from [Hugging Face](https://huggingface.co/zai-org/GLM-5.2), but in practice this needs self-hosting infrastructure, not a local machine.</li>
</ul>

![](https://mintcdn.com/usefulai/Ez-pJDkPpztLE7Cr/images/icons/144/deepseek.com.png?fit=max\&auto=format\&n=Ez-pJDkPpztLE7Cr\&q=85\&s=f47aa3b25f028302bb1e83f032950b2d)

## [DeepSeek V4 Pro](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)

DeepSeek

Low-cost open-weight writing

[View on Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)

The value standout here: open weights and near-GLM writing quality at one of the lowest prices on the list.

Score **42%**
Price **\$0.76 / 1M**
License Open weight
Context **1M**

Strengths

<ul>
  <li>Remarkably cheap for its quality, with open weights and a large context.</li>
  <li>It writes at roughly the level of pricier open-weight rivals while costing a fraction, which makes it our best value pick when you're generating writing at scale and watching cost.</li>
</ul>

Tradeoffs

<ul>
  <li>It's mid-pack on prose, so it's a value play, not a quality leader - the top Claude and Gemini models write clearly better.</li>
  <li>Open weights need self-hosting infrastructure rather than a laptop, and the score reflects its slower thinking mode.</li>
</ul>

How to access

<ul>
  <li>**App** — Available in [DeepSeek Chat](https://chat.deepseek.com/).</li>
  <li>**API** — Accessible via [DeepSeek API](https://api-docs.deepseek.com/news/news260424).</li>
  <li>**Run locally** — Open weights are available from [Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro), but in practice this needs self-hosting infrastructure, not a local machine.</li>
</ul>

![](https://mintcdn.com/usefulai/52KaILwddzz5TNZ_/images/icons/144/mimo.xiaomi.com.png?fit=max\&auto=format\&n=52KaILwddzz5TNZ_\&q=85\&s=fd6b136b0ae0b52bde1ca67e470b4c0b)

## [MiMo-V2.5-Pro](https://mimo.xiaomi.com/mimo-v2-5-pro/)

Xiaomi

Cheap open-weight drafting

[Visit Xiaomi](https://mimo.xiaomi.com/mimo-v2-5-pro/)

A low-cost open-weight option worth knowing if you want cheap, capable drafting from outside the usual US and Chinese labs.

Score **35%**
Price **\$0.76 / 1M**
License Open weight
Context **1M**

Strengths

<ul>
  <li>Very cheap, with open weights and a large context, and it holds its own against other budget open-weight models for everyday writing.</li>
  <li>A reasonable pick if you're routing high-volume, low-stakes drafting and want to keep costs near the floor.</li>
</ul>

Tradeoffs

<ul>
  <li>It's toward the lower end of this list for quality, so it's a budget workhorse, not a model for polished or high-stakes writing. There's no first-party app, and local use means self-hosting, not a laptop.</li>
  <li>For a little more money, DeepSeek V4 Pro writes better.</li>
</ul>

How to access

<ul>
  <li>**API** — Accessible via [OpenRouter](https://openrouter.ai/models).</li>
  <li>**Run locally** — Open weights are available from [Hugging Face](https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro), but in practice this needs self-hosting infrastructure, not a local machine.</li>
</ul>

![](https://mintcdn.com/usefulai/tO2qspLJNjFc61Zv/images/icons/144/anthropic.com.png?fit=max\&auto=format\&n=tO2qspLJNjFc61Zv\&q=85\&s=2077996fd7746bfe8ee85acdd8018de3)

## [Claude Sonnet 5](https://www.anthropic.com/news/claude-sonnet-5)

Anthropic

Everyday Claude writing value

[Visit Anthropic](https://www.anthropic.com/news/claude-sonnet-5)

The sensible everyday Claude for writing - clearly cheaper than Fable or Opus, and good enough for most drafting and editing.

Score **32%**
Price **\$8.00 / 1M**
License Proprietary
Context **1M**

Strengths

<ul>
  <li>A fast, affordable Claude that handles the bulk of routine writing well - drafts, edits, summaries, and clean structure - with the reliability and tone control Claude is known for.</li>
  <li>It's the right default when Fable 5 and the Opus models are more than the task needs.</li>
</ul>

Tradeoffs

<ul>
  <li>It scores below the older Opus rows and the Gemini leaders, so for your most demanding creative or long-form work, step up to Opus 4.6 or Fable 5.</li>
  <li>Think of it as the workhorse, not the showpiece.</li>
</ul>

How to access

<ul>
  <li>**App** — Available in [Claude](https://claude.ai/).</li>
  <li>**API** — Accessible via [Anthropic API](https://docs.anthropic.com/en/api/overview).</li>
</ul>

![](https://mintcdn.com/usefulai/52KaILwddzz5TNZ_/images/icons/144/kimi.com.png?fit=max\&auto=format\&n=52KaILwddzz5TNZ_\&q=85\&s=8d6772935f3f8299f61c9820fd6d176d)

## [Kimi K2.6](https://platform.kimi.ai/docs/guide/kimi-k2-6-quickstart)

Moonshot

Open-weight prose and critique

[Visit Moonshot AI](https://platform.kimi.ai/docs/guide/kimi-k2-6-quickstart)

A capable open-weight writer with a loyal following, good for drafting and sharp critique if you don't need frontier prose.

Score **31%**
Price **\$3.24 / 1M**
License Open weight
Context **262K**

Strengths

<ul>
  <li>A well-liked open-weight model that writes cleanly and is especially handy for editing and critiquing existing text.</li>
  <li>Open weights give you routing and privacy flexibility, and it's a practical, mid-priced option for teams that want an open model for everyday writing.</li>
</ul>

Tradeoffs

<ul>
  <li>Its context window is the smallest here alongside Gemma, which limits very long documents, and it sits low on prose quality. Despite open weights, it's too large for a personal machine, so you're on a hosted API.</li>
  <li>For cheaper open-weight value, DeepSeek V4 Pro wins.</li>
</ul>

How to access

<ul>
  <li>**App** — Available in [Kimi](https://kimi.ai/).</li>
  <li>**API** — Accessible via [Kimi API](https://platform.kimi.ai/docs/overview).</li>
  <li>**Run locally** — Open weights are available from [Hugging Face](https://huggingface.co/moonshotai/Kimi-K2.6), but in practice this needs self-hosting infrastructure, not a local machine.</li>
</ul>

![](https://mintcdn.com/usefulai/Ez-pJDkPpztLE7Cr/images/icons/144/google.com.png?fit=max\&auto=format\&n=Ez-pJDkPpztLE7Cr\&q=85\&s=11e0c725f841c45443116a184b63beac)

## [Gemma 4 31B](https://huggingface.co/google/gemma-4-31B-it)

Google

Laptop-friendly local writing

[View on Hugging Face](https://huggingface.co/google/gemma-4-31B-it)

The one model here you can genuinely run on a normal computer, trading top quality for offline, private, zero-cost writing.

Score **20%**
Price **\$0.34 / 1M**
License Open weight
Context **262K**

Strengths

<ul>
  <li>The most local-friendly pick by far: it runs on a typical machine, so you get offline use, privacy, and no per-token cost.</li>
  <li>Great for private drafting, learning, and low-stakes writing where you want full control and nothing leaving your device.</li>
</ul>

Tradeoffs

<ul>
  <li>It has the lowest prose score here, so expect noticeably weaker writing than any hosted frontier model - fine for notes and casual drafts, not polished work.</li>
  <li>If you can use the cloud at all, almost everything above it writes better.</li>
</ul>

How to access

<ul>
  <li>**API** — Accessible via [OpenRouter](https://openrouter.ai/models).</li>
  <li>**Run locally** — You can run it locally with [Ollama](https://ollama.com/library/gemma4) and [LM Studio](https://lmstudio.ai/models/google/gemma-4-31b) after downloading weights from [Hugging Face](https://huggingface.co/google/gemma-4-31B).</li>
</ul>

***

## How to Choose

When choosing between these models, consider:

* **Access:** Decide first whether you'll use a model in an app, call it through an API, or self-host, because that choice drives cost, privacy, latency, and setup. Most models here are app-and-API; only Gemma 4 31B runs comfortably on a normal machine, and Muse Spark is app-only.
* **Quality:** Our score normalizes the Arena Text Creative Writing Elo, a human-preference ranking of creative prose. It captures voice and style well, but it doesn't measure factual accuracy, SEO structure, or editing reliability - so treat it as a prose signal, not a verdict on every kind of writing.
* **Price:** We use blended cost per million tokens at a 1:3 input-to-output ratio so you can compare on one number. Open-weight models can be cheaper still if you self-host, but only Gemma runs locally without server-grade hardware.
* **Context window:** A bigger window matters when you draft from long sources or many documents at once. Most models here reach about 1M tokens; Kimi K2.6 and Gemma 4 31B are the notable smaller exceptions at 262K.

***

## Other Models We Considered

![](https://mintcdn.com/usefulai/82PG1Up2qz4DPkMj/images/icons/48/anthropic.com.png?fit=max\&auto=format\&n=82PG1Up2qz4DPkMj\&q=85\&s=f84edebba0edca3b7e75a740d0fc318e)[Claude Opus 4.7](https://www.anthropic.com/news/claude-opus-4-7) (Anthropic) — A strong Opus bridge, but 4.6 and 4.8 are the better current picks.

![](https://mintcdn.com/usefulai/82PG1Up2qz4DPkMj/images/icons/48/google.com.png?fit=max\&auto=format\&n=82PG1Up2qz4DPkMj\&q=85\&s=d44aa6f953cf72e889c25d7f615750e5)[Gemini 3 Pro](https://ai.google.dev/gemini-api/docs/models/gemini-3-pro-preview) (Google) — Scored high, but Google shut down the preview, so it's no longer available.

![](https://mintcdn.com/usefulai/82PG1Up2qz4DPkMj/images/icons/48/anthropic.com.png?fit=max\&auto=format\&n=82PG1Up2qz4DPkMj\&q=85\&s=f84edebba0edca3b7e75a740d0fc318e)[Claude Sonnet 4.6](https://www.anthropic.com/news/claude-sonnet-4-6) (Anthropic) — Still widely searched, but Sonnet 5 is the better value now.

![](https://mintcdn.com/usefulai/nqtrSJ8E-k7bZERT/images/icons/48/z.ai.png?fit=max\&auto=format\&n=nqtrSJ8E-k7bZERT\&q=85\&s=f1e232f5604980295eef4f93882dc07f)[GLM-5.1](https://huggingface.co/zai-org/GLM-5.1) (Z.ai) — Writes better than GLM-5.2, but it's the older, less-supported release.

![](https://mintcdn.com/usefulai/C5xOaAf4Os-Vu41o/images/icons/48/openai.com.png?fit=max\&auto=format\&n=C5xOaAf4Os-Vu41o\&q=85\&s=21ac965dc6ee5751127e6d435043629b)[GPT-5.4](https://developers.openai.com/api/docs/models/gpt-5.4) (OpenAI) — A capable prior OpenAI writer, now behind GPT-5.5.

![](https://mintcdn.com/usefulai/C5xOaAf4Os-Vu41o/images/icons/48/qwen.ai.png?fit=max\&auto=format\&n=C5xOaAf4Os-Vu41o\&q=85\&s=dd7a3821c277e432380635e28c70dc83)[Qwen3.5 397B A17B](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) (Alibaba) — A big open-weight Qwen, but weaker at writing than newer picks.

![](https://mintcdn.com/usefulai/C5xOaAf4Os-Vu41o/images/icons/48/openai.com.png?fit=max\&auto=format\&n=C5xOaAf4Os-Vu41o\&q=85\&s=21ac965dc6ee5751127e6d435043629b)[GPT-4.5](https://openai.com/index/introducing-gpt-4-5/) (OpenAI) — A landmark writing model, but that exact version is no longer offered.

![](https://mintcdn.com/usefulai/C5xOaAf4Os-Vu41o/images/icons/48/openai.com.png?fit=max\&auto=format\&n=C5xOaAf4Os-Vu41o\&q=85\&s=21ac965dc6ee5751127e6d435043629b)[ChatGPT-4o](https://platform.openai.com/docs/models/gpt-4o) (OpenAI) — The old ChatGPT default many still expect; that snapshot is gone.

![](https://mintcdn.com/usefulai/82PG1Up2qz4DPkMj/images/icons/48/google.com.png?fit=max\&auto=format\&n=82PG1Up2qz4DPkMj\&q=85\&s=d44aa6f953cf72e889c25d7f615750e5)[Gemini 2.5 Pro](https://ai.google.dev/gemini-api/docs/models/gemini-2.5-pro) (Google) — A familiar baseline, now behind newer Gemini models.

![](https://mintcdn.com/usefulai/C5xOaAf4Os-Vu41o/images/icons/48/deepseek.com.png?fit=max\&auto=format\&n=C5xOaAf4Os-Vu41o\&q=85\&s=b569655af7e98e3467f2aa4cea601b70)[DeepSeek V4 Flash](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash) (DeepSeek) — Cheaper than V4 Pro, but a clear step down in quality.

![](https://mintcdn.com/usefulai/C5xOaAf4Os-Vu41o/images/icons/48/x.ai.png?fit=max\&auto=format\&n=C5xOaAf4Os-Vu41o\&q=85\&s=cc12d6cad1f15f7fed50baa7179965d3)[Grok 4.3](https://docs.x.ai/developers/models/grok-4.3) (xAI) — The newer Grok, but it writes worse than Grok 4.20 here.

***

## Frequently Asked Questions

> [accordion] What is the best LLM for writing right now?
>
> For pure prose quality, Claude Fable 5 is the top of our list. But Claude Opus 4.6 writes almost as well for half the price, so it's the one most serious writers should reach for first.

> [accordion] What is the best LLM for writing for most people?
>
> Claude Sonnet 5 and Gemini 3.5 Flash. Both are strong, affordable, and easy to access in a mainstream app, and they cover the everyday drafting and editing most people actually do.

> [accordion] What is the best free LLM for writing?
>
> Muse Spark, if you're happy working inside the Meta AI app. If you'd rather run something yourself for free, Gemma 4 31B is the only pick here that runs on a normal computer at no per-token cost.

> [accordion] What is the best open-weight LLM for writing?
>
> GLM-5.2 is the strongest current open-weight writer with active support. DeepSeek V4 Pro is the value choice, and Gemma 4 31B is the one you can actually run locally.

> [accordion] What is the best LLM for writing you can run locally?
>
> Gemma 4 31B. It's the only model on this list that runs comfortably on a typical machine. The other open-weight models - GLM-5.2, DeepSeek V4 Pro, MiMo-V2.5-Pro, Kimi K2.6 - technically have downloadable weights but need server-grade hardware.

> [accordion] Why does an older model rank above the newer Claude Opus 4.8?
>
> Because writing quality and version number don't move together. Opus 4.6 was tuned in a way that produces better creative prose than Opus 4.8, even though 4.8 is the newer, stronger all-round model. For writing specifically, 4.6 wins.

> [accordion] Do writing benchmarks match real-world use?
>
> Partly. Our score comes from human preference on creative prose, so it tracks voice and style well. It says little about factual accuracy, SEO structure, or reliable editing, so a high score is a good starting signal, not a guarantee for your exact task.

> [accordion] Is Claude Fable 5 worth the premium over Opus 4.6?
>
> Only for the hardest creative work - fiction, distinctive brand voice, or pieces where prose quality is the whole point. For most writing, Opus 4.6 gets you nearly the same result for far less.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.