icanwrite.app

ChatGPT proofreading, tested

ChatGPT Proofreading: How Many Mistakes It Catches (Tested)

We hid 90 mistakes in a 6,000-word short story and asked the 5 GPT models behind ChatGPT to find them. Here's what each caught, and every mistake the best one missed

Updated Nov 19, 2026

  • Best GPT modelGPT 6.2 Astra89/90 caught · $0.60
  • Everything in one prompt85/90 caughtGPT 6.2 Astra, all 4 checks at once
  • Among all AI modelsTied for 1stOf 22 models, with Claude Fable 5.2

Can ChatGPT proofread?

Yes. The best of the GPT models behind ChatGPT, GPT 6.2 Astra, caught 89 of 90 planted mistakes, tied for the most of the 22 models we tested, with Claude Fable 5.2. Asked to find everything in one prompt, as in a plain “proofread this”, it caught 85. The cheaper GPT-6.1 Sol caught 87 for $0.20, against $0.60.

  • Grammar & spelling
  • Repetitions
  • Redundancy
  • Consistency
  • Combined: all four checks in one prompt
Issues caughtScoreTimeCost
  1. GPT 6.2 AstraOpenAI · Nov 19, 2026
    89/9099%1m 1s$0.60
  2. GPT 6.2 Astra (combined)
    85/9094%1m 44s$0.30
  3. Claude Fable 5.2Anthropic · Nov 10, 2026
    89/9099%4m 8s$2.29
  4. Claude Fable 5.2 (combined)
    88/9098%3m 45s$1.00
  5. Gemini 4 NeonGoogle · Nov 19, 2026
    88/9098%1m 39s$0.70
  6. Gemini 4 Neon (combined)
    87/9097%2m 16s$0.32
  7. GPT-6.1 SolOpenAI · Oct 27, 2026
    87/9097%1m 4s$0.20
  8. GPT-6.1 Sol (combined)
    80/9089%1m 9s$0.08
  9. GPT-6.1 TerraOpenAI · Oct 27, 2026
    79/9088%29s$0.07
  10. GPT-6.1 Terra (combined)
    76/9084%35s$0.02
  11. GPT 5.5OpenAI · Oct 28, 2026
    75/9083%1m 50s$0.94
  12. GPT 5.5 (combined)
    67/9074%1m 50s$0.28
  13. GPT-6 LunaOpenAI · Oct 27, 2026
    59/9066%1m 20s$0.02
  14. GPT-6 Luna (combined)
    60/9067%1m 49s$0.0049

The GPT models, with the best Claude and Gemini for comparison. Click a model to read every comment it left on the story

What ChatGPT catches, by kind of mistake

GPT 6.2 Astra's score on each of the 4 checks, with the other GPT models under it

  • Grammar & spelling44/45Agreement, verb forms, pronouns, real words in the wrong place
    • GPT-6.1 Sol43
    • GPT-6.1 Terra42
    • GPT 5.540
    • GPT-6 Luna35
  • Repetitions15/15The same word twice, a line or chapters apart
    • GPT-6.1 Sol14
    • GPT-6.1 Terra13
    • GPT 5.510
    • GPT-6 Luna11
  • Redundancy15/15Phrases that say one thing twice, like “return back”
    • GPT-6.1 Sol15
    • GPT-6.1 Terra11
    • GPT 5.511
    • GPT-6 Luna5
  • Consistency15/15Details that change: a name, a date, an eye color
    • GPT-6.1 Sol15
    • GPT-6.1 Terra13
    • GPT 5.514
    • GPT-6 Luna8
ModelCaughtGrammar & spellingRepetitionsRedundancyConsistencyAll in oneCommentsExtra flagsTimeCost
GPT 6.2 Astra89/9044/4515/1515/1515/1585/90−4134241m 1s$0.60
GPT-6.1 Sol87/9043/4514/1515/1515/1580/90−79941m 4s$0.20
GPT-6.1 Terra79/9042/4513/1511/1513/1576/90−389729s$0.07
GPT 5.575/9040/4510/1511/1514/1567/90−89961m 50s$0.94
GPT-6 Luna59/9035/4511/155/158/1560/90+16731m 20s$0.02

Free to reuse with credit to icanwrite, under CC BY 4.0

The mistakes ChatGPT missed

GPT 6.2 Astra missed 1 with 4 separate checks and 5 in one prompt. 1 of them no GPT model caught at all.

  1. Grammar & spelling“The survey from twelve years ago, which an engineer from the county”Fragment: the survey has no verbNo GPT model caught itSeparate checksMissedOne promptMissedCaught by 2 of 22 models
  2. Repetitions“a gleaming pair of shoes” … “its point gleaming”Distinctive, chapters apart: gleamingSeparate checksCaughtOne promptMissedCaught by 9 of 22 models
  3. Repetitions“a short, brittle sound” … “tearing the brittle paper”Distinctive, chapters apart: brittleSeparate checksCaughtOne promptMissedCaught by 11 of 22 models
  4. Repetitions“her hands remembered” … “more than her memory did”Close: remembered … memorySeparate checksCaughtOne promptMissedCaught by 5 of 22 models
  5. Redundancy“the week we first began”began is firstSeparate checksCaughtOne promptMissedCaught by 8 of 22 models

One prompt or four?

4 of 5 GPT models caught fewer when one prompt asked for all 4 kinds of mistake than with a prompt for each

  • GPT 6.2 Astra894 checks85One prompt−4
  • GPT-6.1 Sol874 checks80One prompt−7
  • GPT-6.1 Terra794 checks76One prompt−3
  • GPT 5.5754 checks67One prompt−8
  • GPT-6 Luna594 checks60One prompt+1

How to proofread with ChatGPT

Ask for one kind of mistake at a time

GPT 6.2 Astra caught 89 with a prompt for each kind of mistake and 85 with one prompt asking for all of them. Run grammar and spelling first, then repeated words, redundant phrases and details that contradict each other, each in its own request.

Paste the whole chapter

A contradiction only shows when both sides are in the same request. Reading the whole 6,000-word story at once, GPT 6.2 Astra caught 15 of 15 planted contradictions, some of them chapters apart. Don't paste much more than 6,000 words at once, though: that's the length we tested, and on longer text a model can start to miss mistakes.

Ask for quotes, not a rewrite

A rewritten chapter hides what changed, and it changes more than you'd think: editGPT's developer puts it at about 10% of the text for “only fix grammar”, 20–30% for “proofread this” and up to 60% for “improve the flow” (Windows Central). A list keeps your voice: take one fix, leave the next.

Use the strongest model you can

GPT 6.2 Astra caught 89, GPT-6 Luna 59: a wider gap than any prompt made. Which model the ChatGPT app uses depends on your plan, so pick the strongest one it offers for a final pass.

Read every flag before you accept it

Besides the mistakes it caught, GPT 6.2 Astra left 24 comments on text with nothing planted in it, out of 134. A looser prompt flags far more: in one blogger's test, about 11 of ChatGPT's 14 flags on a single post were false alarms. The call is yours.

Treat it as a first pass

Even the best GPT model missed 1 of 90, and 5 in one prompt. Run it before a human proofreader, not instead of one: a person still reads for what no checklist covers.

Before you paste your manuscript

Turn off training on your chats

ChatGPT can use your conversations to train OpenAI's models unless you switch it off: Settings → Data Controls → “Improve the model for everyone”. Your chats stay in your history (OpenAI Help).

No KDP disclosure for proofreading

Amazon KDP asks you to disclose AI-generated content, not AI-assisted: using AI to edit or error-check writing you wrote yourself needn't be reported (KDP content guidelines). Schools and journals set their own rules, so check theirs.

A chapter at a time, not the whole book

Each model read the 6,000-word story, about a chapter, in one request. A whole book won't fit in one, and a chapter is enough for a contradiction to have both its sides in view.

How we tested ChatGPT

One story, 90 planted mistakes

An original 6,000-word short story, clean apart from what we planted: 45 grammar and spelling errors, 15 repeated words, 15 redundancies and 15 contradictions. The hard kind, not typos a spellchecker would catch.

Through OpenAI's API, not the app

We called each GPT model through OpenAI's API with the same frozen prompts every model in the benchmark gets. The ChatGPT app adds instructions of its own, and which model it uses depends on your plan, so its answers can differ from these.

What counts as a catch

A mistake is caught when one of the model's comments lands on it. Scoring is plain text matching: no AI judges another AI. Each model reads the whole story once per check, and once more with all 4 checks in a single prompt, scored on its own.

The full benchmark · Claude, Gemini and moreBest AI for Proofreading and Copy Editing: 22 Models TestedSee every model

Free sample

Want these checks on your own draft?

The same 4 checks, plus 15 more — from dialogue to point of view. Runs on GPT-6.1 Sol, which caught 87 of 90 mistakes here. See all 19 checks

  • No card needed

Drag & drop your manuscriptUpload your manuscript

.docx, .epub or .md

or
Choose file

Questions about ChatGPT proofreading

Yes, on a strong model. The best GPT model we tested, GPT 6.2 Astra, caught 89 of 90 mistakes planted in a 6,000-word story, tied for first of the 22 models in our benchmark with Claude Fable 5.2. GPT-6.1 Sol caught 87. The weakest, GPT-6 Luna, caught 59, so which model answers matters more than how you ask.

See every comment GPT 6.2 Astra left →

Very well, though not perfectly. GPT 6.2 Astra caught 44 of 45 planted grammar and spelling errors, GPT-6.1 Sol 43, GPT-6.1 Terra 42, GPT 5.5 40 and GPT-6 Luna 35. No GPT model caught a sentence fragment with no verb, and only Astra caught “longer than any shop on Wharf Street” (it should be any other shop). None of the 45 were typos a spellchecker would catch.

Ask for one kind of mistake at a time, and ask it to quote each one instead of rewriting the text. GPT 6.2 Astra caught 89 of 90 with four separate prompts (grammar and spelling, repetitions, redundancy, consistency) but 85 when one prompt asked for all four. GPT-6.1 Sol fell from 87 to 80, GPT 5.5 from 75 to 67.

At the top it's a tie, and GPT costs less. GPT 6.2 Astra and Claude Fable 5.2 both caught 89 of 90: Astra for $0.60 in about a minute, Fable for $2.29 in four. Claude Opus 5.5 caught 88 for $0.74, and GPT-6.1 Sol 87 for $0.20.

Yes, if it sees the whole story. GPT 6.2 Astra and GPT-6.1 Sol caught all 15 contradictions planted in it, from a grandmother called Ruth and then Rose to a Saturday that's wrong once you count the days. GPT 5.5 caught 14, GPT-6.1 Terra 13 and GPT-6 Luna 8. A contradiction only shows when both sides are in the same request, so paste the whole chapter, not a paragraph.

GPT 6.2 Astra: 89 of 90 for $0.60. GPT-6.1 Sol caught 87 for $0.20, the better value, GPT-6.1 Terra 79 for $0.07, GPT 5.5 75 for $0.94 and GPT-6 Luna 59 for two cents. We tested them through OpenAI's API. Which model the ChatGPT app uses depends on your plan.

Not in one go: work a chapter at a time. Our test story is about a chapter, 6,000 words, and every model read it in one request. A chapter also keeps both sides of most contradictions in view, which is how GPT 6.2 Astra and GPT-6.1 Sol caught all 15 in the story. For names and dates that change across a whole book, keep a list of them and check each chapter against it.

It will if you ask it to proofread and let it return the text. Ask for a list of mistakes instead, quoting the exact words, and nothing changes until you make the fix. editGPT's developer puts the change at about 10% of the text for "only fix grammar", 20–30% for "proofread this" and up to 60% when asked to improve the flow.

ChatGPT can use your chats to train OpenAI's models unless you turn it off: Settings, then Data Controls, then switch off "Improve the model for everyone". Your chats stay in your history, and it works when you're signed out too.

OpenAI's help page on training →

No. KDP asks you to disclose AI-generated content, meaning text, images or translations an AI created, but not AI-assisted content: using AI to edit, refine or error-check writing you created yourself doesn't need to be reported. Schools and journals set their own rules, so check theirs.

KDP's content guidelines →

We didn't test human proofreaders, so we can't put a number on it. The best GPT model, GPT 6.2 Astra, missed 1 of 90 planted mistakes with separate checks and 5 with one prompt, among them a sentence fragment no GPT model caught. Use it as a first pass before a human, who also reads for what no checklist covers.

It depends on the model. Apart from the 90 planted mistakes the story is clean. GPT-6.1 Sol left only 4 comments that landed on none of them and GPT-6.1 Terra 7, while GPT 6.2 Astra left 24, the most of any model we tested, and Claude Fable 5.2 17. Some of those are fair style notes, not errors, so read each before you accept it.

These checks come from icanwrite, a markdown editor for writers