The best AI model for proofreading (we tested them all)

Result snapshot updated September 29, 2026. GPT-5.6 Sol tops the benchmark at 96.5% with medium reasoning. The six selected models below all land within two points of each other, while their cost per benchmark run ranges from $0.41 to $1.88. For most proofreading, the last point of accuracy isn’t worth paying three times as much for, so our recommendations lead with value:

• Best value overall: GPT-6 Sol. With no reasoning it scores 95.4% for $0.41 per run - about one point behind GPT-5.6 Sol at a third of the price, fixing 109 errors per dollar to Sol’s 37. It’s also fast, at 38 fixes per minute.
• Best Claude: Opus 5.5. With high reasoning it scores 94.7% for $0.46 per run - one point behind Claude Opus 5 (95.7%) at half the price, or about twice as many fixes per dollar. Its 0.4% seeded-error bad-fix rate measures incorrect repairs of seeded issues, not overall document safety. The trade-off is completeness - it leaves 4.9% of errors untouched, more than the other flagships.
• Only if every point matters: GPT-6 Astra. It scores 96.1% with low reasoning, within half a point of GPT-5.6 Sol, for $1.13 per run - and it fixes errors nearly five times as fast (40.3 fixes per minute to Sol’s 8.4). GPT-5.6 Sol itself is the most accurate model, missing just 2.3% of errors, but costs more than Astra and runs about five times slower.

Cost per run is the dataset-weighted average cost of one source-text evaluation, including the proofreading model and fallback judge. It is not the price of processing the whole corpus or a Revise subscription price.

Some other noteworthy mentions based on the full data:

• Opus 5 likes its reasoning dial low; Opus 5.5 does not. Opus 5 scores best with low reasoning (its no-reasoning variant drops all the way to 86.5%), and the older Opus models (4.6, 4.7, 4.8) do best with no reasoning at all. Opus 5.5 reverses the pattern: it scores highest with high reasoning and drops to 91.6% on low.
• Gemini 3.7 Flash is the new budget pass - it scores 91.4% for about $0.12 per run while fixing 31.4 errors per minute.
• GLM 5.3 Flash is the cheapest model to crack 90%, scoring 90.1% for about $0.07 per run.
• Grok 4.5 remains xAI’s best at 93.4%, but GPT-6 Sol now beats it on quality, price, and speed. The newer Grok 4.6 scores slightly lower (92.9%) at a higher price and runs at about 4 fixes per minute.
• There are certain models it just never makes sense to use, such as Claude Sonnet (inferior to Gemini and GLM) and Grok 4.2 - beaten on both quality and price by xAI’s own 4.3 and 4.5.

Here are six selected models and their best-scoring reasoning configurations. The complete ranking includes additional models. See the full results for more data and scatterplots.

A proofreading benchmark only means something if the test is realistic. Here’s the short version of how ErrataBench works:

• Real text, real errors. We start from genuine source documents across 11 datasets - novels, legal text, scientific writing, nearly 99,000 words in all - and use a model pair to seed small, realistic mistakes drawn from a taxonomy of error categories, with a second model reviewing each one before it’s accepted.
• No cheat sheet. Each model is asked to proofread carefully, but is never told what kinds of errors are hidden or how many there are - just like a real editing job.
• The same tools for everyone. Models work through a simple agent loop with find-and-replace tools, get the same text chunks and the same number of turns. Canonical corrections are scored automatically. A fallback LLM judge evaluates alternative repairs of seeded issues; this result set records Claude Opus 4.6 with low reasoning as the judge.

Every model gets the identical prompt, tools, and text - so the only configurations include the model and its reasoning setting. Full methodology, source code, and raw results are published and reproducible.

Beyond the headline fix rate, two numbers tracked in the full results matter just as much: missed errors - the ones a model never touched (a completeness problem) - and bad fixes, the repairs of seeded issues it attempted that were actually wrong. The second is the dangerous one - a model that “corrects” text incorrectly makes your writing worse. The best models keep bad fixes near 1%.

The most accurate model isn’t always the one you should reach for. Proofreading is often high-volume - a long report, a whole manuscript, a stack of documents - so cost and speed matter.

At the very top, GPT-5.6 Sol leads on quality at about $1.27 per benchmark run, and GPT-6 Astra comes within half a point for $1.13. Those last points are expensive, though. GPT-6 Sol reaches 95.4% for about $0.41 with no reasoning - cheaper than every model that beats it, and far cheaper than GPT-5.5 ($1.88), which scores lower. Claude Opus 5.5 makes the same trade on the Claude side: 94.7% for about $0.46, against Opus 5’s 95.7% for $0.96. Between the two, GPT-6 Sol edges Opus 5.5 on quality, price, and speed, while Opus 5.5 makes fewer bad fixes.

Below the flagships, Gemini 3.1 Pro reaches 92.1% for roughly $0.37, Gemini 3.7 Flash reaches 91.4% for about $0.12, and GPT-6 Luna reaches 90.9% for about $0.15. Kimi K3 matches Gemini 3.1 Pro’s 92.1%, but at about $0.91 per run and a much slower pace, Gemini is the better buy at that quality level.

That’s the whole argument for not marrying a single model: use a fast, cheap one for quick passes and escalate to a flagship for the writing that really matters. You can explore the full accuracy-vs-cost tradeoff - plus speed and consistency - on the interactive ErrataBench scatterplot.

No model is perfect, and the errors they miss aren’t random. Aggregated across every model we tested, the easiest category to catch is spelling & word form - the obvious typos a basic checker would flag too, resolved about 92% of the time. The hardest is meaning, flow & style, resolved only around 82% of the time - the subtle, judgment-heavy issues where reasonable editors can disagree. The full results break performance down by error category.

The practical lesson: AI proofreading is excellent, but it’s a collaborator, not an oracle. You still want to see and approve what it changes - which is why the how of AI proofreading matters as much as which model.

Picking a model is half the battle. The other half is the tool you run it in. Pasting paragraphs into a chatbot and copying results back is slow, loses your formatting, and - worst of all - gives you no easy way to see exactly what changed.

Revise is an AI document editor designed for exactly this:

• Every top model, one click away. Because Revise is multi-provider, you can proofread with GPT-6 Astra, run a cheaper pass with GPT-6 Sol or Claude Opus 5.5, and get a second opinion from Gemini 3.1 Pro - all in the same document, no accounts or API keys to juggle.
• Whole-document context. The AI reads your entire draft, so it catches inconsistencies a paragraph-at-a-time chatbot never would - a term spelled two ways, a tense that drifts.
• Tracked changes you control. Every fix appears as a red/green diff you accept or reject one by one. The AI proposes; you decide. Nothing is overwritten silently.
• Full revision history. Scrub back through every version to see exactly how the document evolved and which edits came from you versus the AI.

In other words, Revise turns “which model is best” from a one-time bet into a dial you can turn per task - with the review controls that make it safe to let AI touch a document that matters.

What’s the best AI model for proofreading?

For most people, GPT-6 Sol with no reasoning: it scores 95.4%, about one point behind the leader, for a third of the price. If you prefer Claude, Claude Opus 5.5 with high reasoning scores 94.7% at half the price of Opus 5 and makes the fewest bad fixes of any top model. The most accurate model overall is GPT-5.6 Sol at 96.5%, with GPT-6 Astra close behind at 96.1% and nearly five times faster.

Is ChatGPT or Claude better at it?

In the selected configurations, GPT-5.6 Sol beats Claude Opus 5 by 0.8 percentage points. GPT-6 Astra also outscores Opus 5 while fixing 40.3 errors per minute to its 22.5, though Opus 5 is a little cheaper per run ($0.96 vs. $1.13). Claude Opus 5.5 has a 0.4% seeded-error bad-fix rate. This measures incorrect repairs of seeded issues, not overall document safety.

How accurate is AI proofreading, really?

The leading configurations repair more than 90% of the benchmark’s seeded errors. This result does not measure every unwanted change elsewhere in a document, so review the proposed edits before accepting them.

And since the models keep leaping ahead of each other, the smartest move isn’t to pick one forever - it’s to use a tool that lets you switch. Try it on your own writing: paste a document and proofread it with the top AI models. It’s free to use with basic AI, and with a paid upgrade it offers the most powerful AI features of any word processor. And if you want to go deeper on the data, the full ErrataBench results are open and interactive.

Quoting this post? All the data here is free to reuse with a link back. Copy a ready-made citation.