Same Essay, New AI Score? GPTZero and Copyleaks Changed Models in September 2026, and Pangram Set a Deadline

Possibly because the detector changed, not your essay. GPTZero's release notes say its new model, which it calls 4o, was released on 18 September 2026, and its announcement says it became the default for all users on 20 September. Copyleaks released version 11.0 of its AI text detection model on 22 September. Pangram said its older model, 3.3.2, would stay the default "until September 30, 2026". If your two scans fall on either side of one of those dates, the gap between them may have nothing to do with what you did to the text, whichever way the number went.

This article is about those three third-party detectors. If the two numbers you are comparing came from Turnitin, its model runs on its own release schedule, covered in Why Did My AI Score Change When I Resubmitted the Same Paper?

None of these dates, on its own, tells you which model read your scan. Each announcement leaves a gap, and the gaps are different.

The release-notes entry is headed "September 13, 2026 (2026-09-13-base, Model 4.10b aka 4o) | released on September 18, 2026". So the entry is dated 13 September, but the release day it gives is the 18th. It lists two improvements: "Decreased false positive rate on a variety of domains" and "Improved performance on third party benchmarks".

GPTZero's post introducing the model is dated 24 September and gives a different day: "As of 20 September 2026, the updated model is available to all users and is the default model offered on GPTZero."

We are not going to decide which date is right. A scan from before 18 September predates both. A scan from 20 September onward comes after both. A scan in between could have gone either way.

GPTZero changes models often. Its release notes list eleven model entries dated 2026, ten "base" models and one multilingual, so two scans a couple of months apart may well have been read by different models even without this one.

One thing you may notice on a newer result: gray text. The 4o post presents it as new ("We've also introduced a new highlighting color: gray") and says "Gray spans show which pieces of text are excluded by our AI detector", which for now means headings. The release notes put the change underneath it, leaving headings out of the model's input, earlier: 1 August (Model 4.8b), "headers are now excluded from our model's input". For comparing scores, that earlier date is the one that matters. In a GPTZero scan from before 1 August your headings counted; in a later one they did not.

On 22 September, Copyleaks' changelog says it has "released version 11.0 of the AI Text Detection model", which "improves detection accuracy and extends coverage to recently released large language models".

The next paragraph shows who it is for: "No request changes are needed - the modelVersion field in the response now returns v11.0." This is Copyleaks' API documentation, the service other software is built on. It says the new model is live on the API's AI text detection. It does not say when the Copyleaks website, or the version built into your school's systems, made the same switch.

One university did warn its staff. A post dated 14 September on the blog of the Center for Faculty Development and Innovation at Canisius University said: "On Tuesday, September 22nd, Copyleaks will be updating their AI text detector model to help ensure accuracy with AI content detection." It added that "instructors should monitor their scans and be aware of variances with scan results and adjust settings accordingly." That is one school passing on a notice about a week in advance, and it names no version number. It does suggest the change was expected to show up in instructors' scans and not only in developers' code.

If you need to pin a Copyleaks result to a model, the API has carried the answer since 16 April 2024, when Copyleaks added a `modelVersion` field to its AI detection results. Whether that number appears on the report you were shown depends on the system that produced it.

Pangram 4 was "released through the API" on 29 July 2026. The migration guide then says: "Pangram 3.3.2 is still the default and will remain live until September 30, 2026." The launch post spells out what that means for software that doesn't ask for a specific model: "Default requests that don't specify a model will continue routing to Pangram 3 using the previous billing system, until Pangram 3 is deprecated."

So, by Pangram's account, a platform calling Pangram without naming a model got 3.3.2 through September, and under Pangram's plan would stop getting it after the 30th.

What we could not confirm is that the switch happened. On 2 October, none of the Pangram pages we read said so: the migration guide (last modified 7 August), the launch post, the Pangram 4 model card, and the API reference. The API reference describes a "default" setting that "follows Pangram's current default model" without naming which model that currently is. These are also statements about the API. They don't say which model Pangram's own website used during the changeover.

There is a visible tell, though only in a narrow place. Pangram's migration guide says its per-segment ("window") labels changed: "AI-Assisted replaces the previous Lightly AI-Assisted and Moderately AI-Assisted labels." Pangram 4 results through the API also return a version of "4.0". So if a segment in an earlier result was labelled "Moderately AI-Assisted" and segments in a later one only ever say "AI-Assisted", the two did not come from the same model. Two limits. This is about the labels on segments, not the overall verdict, which Pangram 4's own API example also gives as "AI Assisted". And Pangram's December 2025 announcement says free users saw "lightly AI-assisted text as Human, and moderately AI-assisted text as AI-generated", so on a free account you may never have seen the two old grades at all.

The updates are described in general terms. GPTZero says it decreased false positives "on a variety of domains"; Copyleaks says version 11.0 "improves detection accuracy" and covers newer language models. Neither says which way any one essay will move.

So if your score fell after one of these updates, that is not evidence your rewrite worked. If it rose, that is not evidence your rewrite made things worse. The two numbers mix your edits with a model change, and nothing in either result separates them.

1. Line up the dates. Write down the day of each scan and check it against the dates above. If one of those dates falls in between, treat the two numbers as readings from two different instruments.
2. Check the setting matched. Some tools let the person scanning choose a mode. GPTZero's announcement refers to an "Advanced Scan" mode, and Originality.ai lets you scan by model ("Classic") or with an "AI Allowance" of 0%, 5%, 15%, 25% or 40%. A different mode is a different measurement.
3. To see what your edits did, scan both versions on the same day. Old draft and new draft, same tool, same setting, one after the other. The model is then almost certainly the same for both (if the result shows a model name or version, check that the two match, since a scan during a changeover like 18 to 20 September could go either way), and the difference belongs to the text. It still only tells you what that one tool thinks that day, not what your school's tool will say.
4. Keep the whole result, not the number. Save a screenshot or the report itself with the date, the mode, and any model name or version it shows. A retyped "34%" can't be checked against a release note later.
5. If someone else ran the scan, ask three things: which tool, which day, and which setting or version. Without those, a number cannot be compared with anything, including your own self-check.

Two different tools disagreeing on the same file comes down to other reasons, covered in Why the same text scores differently on every detector. For what the labels on a GPTZero result mean, see GPTZero Mixed and AI-Polished: What the Labels Mean.

Will my old result be updated to the new model? None of the three announcements quoted here says earlier results are re-scored. Treat a saved result as a reading by whatever model ran on that day. Turnitin's model release notes, for comparison, carry the line "This release will not retroactively update previously-generated AI writing reports."

My instructor's scan is from before the update and mine is from after. Which one counts? Neither cancels the other. They are readings by two different models, and quite possibly two different settings. The one your course looks at is the one your instructor ran, so the useful thing to ask is which tool, which day and which setting produced it. A scan you run now, on the newer model, is not a re-run of it.

Does this affect my Turnitin score? No. Turnitin is a separate company with its own model releases. Check its dates the same way.

• GPTZero, GPTZero Release Notes (Model and API), entries for 13 September 2026 (Model 4.10b) and 1 August 2026 (Model 4.8b); read 2 October 2026.
• GPTZero, Introducing GPTZero 4o, dated 24 September 2026; read 2 October 2026.
• Copyleaks, What's New (API documentation), entries for 22 September 2026 and 16 April 2024; read 2 October 2026.
• Canisius University, Center for Faculty Development and Innovation, Copyleaks AI Text Detector Model Update, dated 14 September 2026; read 2 October 2026.
• Pangram, Pangram 4 Migration Guide (6 August 2026) and Introducing Pangram 4 (29 July 2026); Introducing Pangram 3.0 with AI assistance detection (11 December 2025); AI Detection and Models in Pangram's API reference; all read 2 October 2026.
• Originality.ai, Introducing AI Allowance; read 2 October 2026.