Which AI Models Does Turnitin Detect? The Named List, and the Gap

Turnitin's English AI writing FAQ names 32 models, from `GPT-4o` to `Grok-4.1`, and says its detector "can detect content from" them "and tools based on these LLMs as well". Being on that list is not a promise that your text gets caught; the same page says "we might miss some AI written text". Not being on it is not a pass either. Model names that other detector companies have been writing about, including GPT-6 Astra, Claude Opus 5, Gemini 3.7, Grok 4.6 and Fable 5.1, appear in none of Turnitin's 461 English help articles as of 17 September 2026. That means Turnitin has not named them there. The FAQ does not say what happens to text from a model it has not listed; the sentence right after its list says "We will continue to expand our detection capabilities to other models in the future."

So the list answers a narrower question than the one you probably came with. It tells you what Turnitin has put its name to. It does not tell you what score a particular paper will get, for any model, listed or not.

Below: the list as printed, what the wording commits Turnitin to, how we counted the gap, and what to do with that.

It lives in Turnitin's AI writing detection capabilities FAQ. Its first line limits it to English: "The following content and FAQs pertain to our AI writing detection capabilities for English submissions only." Going by its timestamp, it was last edited on 28 August 2026.

The sentence starts "Currently, Turnitin's AI writing detection model for English submissions can detect content from" and then runs through the models, each with a release date in brackets. We split it on commas, from the first model to the words "and tools based on these LLMs", which gives 32 entries. `GPT-5.2` and `GPT-5.2-pro` count as two; the closing "tools based on these LLMs" is not counted as an entry. Spelling and odd spacing are Turnitin's:

We left the release dates out of the table on purpose. The list is printed twice on the same page, once under "How does it work?" and again under "Which AI writing models can Turnitin's technology detect?", and the two copies are identical except for two dates. The first gives `o1-mini` as released 2024-09 and the second as 2024-12. The first dates `Claude Sonnet-4.6` to 2026-02, the second to 2026-05. We are not going to pick one for Turnitin. If a date matters to your argument, quote the copy you mean and say that the other copy disagrees.

One more list gets confused with this one. Turnitin's release notes entry for 18 August 2026 names models too, but it is announcing Arabic detection, and it says so: "our Arabic AI writing model is a separate model from our English AI writing, Spanish AI writing, and Japanese AI writing models." That entry is where Opus 4.6 (printed there as `Claude Opus 4.5 (released 2025-11) and 4.6`), `OpenAI o3` and `Qwen 3 Next` come from. None of those three is on the English list; the only Opus entry there is `Claude-Opus- 4.5`. If you write in English and read that release note as "Turnitin added Opus 4.6", you have read the wrong model's list.

The verb in the English model list is "can detect content from". In that sentence it is not "trained on" and it is not "tested against". The difference is easy to lose, because Turnitin describes its Japanese detection with one of those other verbs: a 2025 release note says the Japanese models were "trained to detect text likely generated with" GPT-4 and GPT-4o. Do not carry the Japanese verb over to the English list.

What the verb does not come with is a number. No entry on the list has a detection rate beside it, and the list does not rank one model as harder to catch than another. So "Claude is detected less often than GPT" is not something you can source to this list.

The same FAQ is also plain about the cost of its low false-positive target:

In order to maintain this low rate of 1% for false positives, there is a chance that we might miss some AI written text in a document. We're comfortable with that since we do not want to incorrectly highlight human-written text as AI-written.

It then gives an example: if it identifies 50% of a document as likely written by an AI tool, the document "could contain as much as 65% AI writing". That applies to listed models too.

And the list is not fixed. Asked whether the AI percentage will change as the detector and the models evolve, the FAQ says "it is likely that our detection capabilities will also change, affecting the AI percentage", and then: "However, for a submitted document, the AI percentage will change only if it's re-submitted again to be processed." So the percentage on a report you already have stays as it is unless the document is resubmitted.

Other detector companies have been publishing posts about model names that are not on Turnitin's list. Originality.ai's blog has "Is GPT-6 Astra Content Detectable?", "Is Gemini 3.7 Flash Content Detectable?" and "Is Grok 4.6 Content Detectable?"; Pangram's has "Can Pangram Detect Claude Opus 5?" and "Does Pangram Detect Fable 5.1?". We are not reporting what those posts found, and we have not checked the models themselves against their makers' announcements. We used the names for one thing: to see whether Turnitin's documentation mentions them.

It does not, as far as its English help center goes. Here is exactly what "does not" means.

• Corpus. Every English article in Turnitin's help center at guides.turnitin.com: 461 articles. We got that number two ways on 17 September 2026 (UTC): by paging the help center's public article API to the end (five pages, 461 unique articles) and from the help center's sitemap (461 English article addresses). The two sets matched with nothing left over on either side. A second pass through the API about half an hour later returned 461 again.
• Method. Tags stripped from each article body, then an exact, case-sensitive text search.
• Searches with zero hits in all 461: `GPT-6`, `Astra`, `Opus 5`, `Opus-5`, `Opus- 5`, `Sonnet 5`, `Gemini 3.7`, `Gemini-3.7`, `Grok 4.6`, `Grok-4.6`, `Fable`.
• Proof the search was working. Same corpus, same search: `Nova-2-lite` shows up in 1 article (the FAQ, twice, once in each copy of the list), and `GPT-5.4` in 2 articles. So the search is live. It matches exact strings, which is why the list above includes the spaced and hyphenated spellings Turnitin itself uses, such as `Opus- 5` after the list's own `Claude-Opus- 4.5`.
• Not covered. turnitin.com itself, meaning the marketing site and Turnitin's blog. We did not search those, so we are not saying Turnitin has never mentioned these names anywhere. The claim stops at the help center.

Now the part that matters more than the count. A missing name tells you what Turnitin has published, not what its detector does. The FAQ does not say what happens to text from a model it has not listed. What it does say about other models looks ahead. In both copies, the sentence right after the list is:

We will continue to expand our detection capabilities to other models in the future.

Its answer on future-proofing for models "yet to emerge" says "we are continuously updating our model to detect these additional LLMs", and two sentences later: "Since then, we've expanded our detection capabilities to include a large number of other LLMs as mentioned in the FAQ earlier." The list itself ends with "and tools based on these LLMs as well" without saying whether a later model from the same maker counts. None of these sentences says an unlisted model is caught today, and none says it is missed. So this article does not say either.

1. Read the list yourself, on the day you need it. Open the FAQ, search the page for "can detect content from", and note the page timestamp. That page was last edited 28 August 2026 when we checked, and it is edited in place. A screenshot with a date on it is worth more in a meeting than any article's paraphrase, ours included.
2. Match the exact model name, not the app name. The list names models (`Claude Sonnet-4.5`, `Gemini 2.5 Flash`), not chat apps. If you cannot tell which model a tool was running, the list cannot settle anything for you, and the "tools based on these LLMs" clause does not name any tool.
3. Check the language. An English paper goes to the English list. Spanish, Japanese and Arabic submissions each have their own model and their own list, and they do not match the English one.
4. Keep the list out of any argument about your own paper. Whichever side of it your model is on, the same FAQ has a sentence that cuts the other way, and neither side is about your document.
5. If you already have a report, work from the report. The percentage on a submitted document only changes when it is resubmitted, so the report you are holding stays as it was scored. How to read a Turnitin AI writing report covers what that number includes. If you wrote the flagged text yourself, preparing a response and keeping version history do more for you than the model list will.

Is Claude Opus 4.6 on Turnitin's list? Not on the English one. The English FAQ's Claude entries stop at `Claude Sonnet-4.6`, `Claude-Haiku-4.5` and `Claude-Opus- 4.5`. Opus 4.6 is named in the 18 August 2026 release note, which describes the Arabic detection model.

Did Turnitin update its English model list in August? The FAQ page was last edited on 28 August 2026, but an edit timestamp does not say which part changed, and we have no complete copy of the list from before that date to compare against. The list in the 18 August release note belongs to the Arabic model.

Does the list tell me how likely my model is to be caught? No. The list has no per-model figures. The figures the FAQ does give are detector-wide, such as the 1% false-positive target and the example that a 50% result "could contain as much as 65% AI writing".

Which release date is right for `Claude Sonnet-4.6`? The page gives both 2026-02 and 2026-05, one in each copy of the list. The Arabic list uses 2026-05. We would quote the model name and leave the date out.

This page is the overview. Each of these goes deeper on one model family or one boundary:

• Does Turnitin detect GPT-5 and other new AI models?
• Does Turnitin detect Claude specifically?
• Does Turnitin detect Gemini and Copilot?
• Does Turnitin detect Claude, Copilot or Gemini? goes into why a coverage list is not a per-model ranking and what Copilot's absence does not show.
• Does Turnitin detect DeepSeek?
• Does Turnitin detect Grok?
• Does Turnitin detect LLaMA?
• Does Turnitin detect Mistral AI text?
• Does Turnitin detect Perplexity AI?
• Turnitin Arabic AI detection: what changed for Arabic writers compares the Arabic list with the English one.
• Can Turnitin detect AI in non-English submissions? covers the separate Spanish, Japanese and Arabic models.

All read logged out on 17 September 2026 (UTC). Turnitin edits these pages in place, so reopen them before relying on any of this.

• Turnitin Guides, Turnitin's AI writing detection capabilities FAQs. Page timestamp 28 August 2026. Source of the model list, both copies, and every FAQ quotation above.
• Turnitin Guides, Turnitin release notes, entry dated 2026 August 18, on Modern Standard Arabic detection. Also the source of the Japanese "trained to detect" wording, in a 2025 entry.
• Turnitin Guides, all 461 English (en-us) help center articles, for the name search, counted through the help center's public article API and its sitemap.
• Blog posts used only as the source of the model names in the section on names that are not on the list: Originality.ai, Is GPT-6 Astra Content Detectable?, Is Gemini 3.7 Flash Content Detectable? and Is Grok 4.6 Content Detectable?; Pangram, Can Pangram Detect Claude Opus 5? and Does Pangram Detect Fable 5.1?.