BenchLM recommendation
Best LLMs for Writing in 2026
As of July 23, 2026, the top model in best llms for writing on the BenchLM leaderboard is MAI-Thinking-1 with a score of 97.8.
Last verified: July 23, 2026
There is no single "writing benchmark," so BenchLM ranks writing capability by the instruction-following category — the best available proxy for whether a model matches your tone, structure, and length constraints — read alongside Arena Elo, the human-preference signal. Claude models hold the top Arena Elo scores; the instruction-following table below shows who follows a brief most reliably.
Unless noted otherwise, ranking surfaces on this page use BenchLM's provisional leaderboard lane rather than the stricter sourced-only verified leaderboard.
Bottom line: instruction following is what separates writing models in practice. Claude Fable 5 holds the top Arena Elo (1508) for human preference; the instruction-following leaders below execute a brief most faithfully.
MAI-Thinking-1 leads this ranking with a score of 97.8, followed by Qwen3.5-27B (93) and Agents-A1 (93). There is meaningful separation between the top models, suggesting genuine performance differences.
The best open-weight option is Qwen3.5-27B (ranked #2 with a score of 93). Open-weight models are highly competitive in this category — self-hosting is a viable alternative to proprietary APIs.
This ranking is based on provisional weighted averages across the scoring benchmarks in instructionFollowing tracked by BenchLM.ai. For detailed model profiles, click any model name below. To compare two specific models head-to-head, use the "vs #" links.
What changed
MAI-Thinking-1 leads the instruction-following category at 93.3.
GPT-5.4 close second at 92.6 — the reliable structured-writing workhorse.
Claude Opus 4.6 strong instruction following (91.3) with Claude's prose quality.
How to choose
Long-form drafts and editing?
Claude Fable 5 — top Arena Elo, strong instruction following
Structured marketing copy?
GPT-5.4 — precise formatting at mid-tier price
Budget content pipelines?
Check price-vs-performance for the cheapest 85+ IF score
The full writing analysis?
The writing deep-dive compares creative-writing Elo directly
Full Rankings (31 models)
Key Takeaways
The top model is MAI-Thinking-1 by Microsoft with a provisional score of 97.8.
The best open-weight model is Qwen3.5-27B at position #2.
31 models are included in this ranking.
Score in Context
What these scores mean
Writing quality has no direct benchmark, so this page ranks by instruction following — whether the model does what the brief asked. Read it with Arena Elo, the blind human-preference score, for prose quality.
Known limitations
Style is subjective and prompt-sensitive. Instruction-following scores reward constraint compliance, not voice — a model can follow your brief perfectly and still write flat prose. Test your actual editing workflow.
Best LLMs for Writing FAQ
What is the best LLM for writing?
Claude Fable 5 is the strongest overall pick — it holds the highest Arena Elo BenchLM tracks (1508), the best human-preference signal for prose, with top-tier instruction following. For structured copy at lower cost, GPT-5.4's 92.6 instruction-following score makes it the workhorse choice.
What is the best LLM for creative writing?
Human preference is the best available signal for creative prose, and Claude models lead it: Claude Fable 5 holds the top Arena Elo on BenchLM's board. Non-reasoning models often feel more natural for iterative drafting because they respond without a thinking pause.
What is the best AI for resume writing?
Resume writing is an instruction-following task — strict format, tight length, specific tone. Any model above ~90 in the table handles it well; the practical differences are price and speed, so a mid-tier model like GPT-5.4 or Claude Sonnet 5 is the sensible default.
Are benchmarks meaningful for writing quality?
Partially. Instruction following measures whether the model obeyed the brief — essential for professional writing — and Arena Elo captures blind human preference. Neither measures your voice. Use the scores to shortlist, then run a 10-prompt bake-off in your own editing workflow.
Explore More
Choose a model with this week’s evidence
Join 2,000+ readers for ranking moves, pricing changes, and the claims that still need proof.
One email each week. Unsubscribe anytime.