Pangram and GPTZero Each Win Their Own Benchmark; Two Independent Tests Found Pangram Level or Ahead

Each benchmark ranks its own publisher first. GPTZero's comparison page, dated 9 October 2025, gives GPTZero 99.6% accuracy, a 0.13% false positive rate and 99.4% recall, against Pangram at 97.5%, 0.20% and 95.4% ¹. Pangram's false positives post, dated 15 September 2026, puts its own rate at roughly 1 in 10,000 and GPTZero's at 2.01% ².

Both pages are specific, and neither can be set against the other. GPTZero's own comparison page puts its rate at 0.13% ¹, while the 1% Pangram measures against is the figure Pangram attributes to GPTZero ². The two companies don't even agree on what GPTZero claims.

GPTZero's table does its real work in the per-model rows. It has GPTZero catching 97.2% of o3 output against Pangram's 85.1%, and 96.6% of Gemini 2.5 Pro output against 85.6% ¹. Both are 2025 readings of a 2025 product.

Pangram 4 replaced Pangram 3.3.2 on 29 July 2026 ⁶, so the Pangram in GPTZero's table isn't the one a district would license this term. A comparison ages as fast as the products in it do, and the October 2025 page carries no later rerun ¹.

Because each was measured on a set the other company can't inspect. GPTZero's post says the table reuses the set from its September 2025 Copyleaks and Originality.AI benchmark ¹, and that benchmark describes 3,000 samples, half human and half AI ⁹. Neither post publishes the set. Pangram's post raises that exact objection about published competitor figures: "a naked number in isolation like this does not tell the whole story" ².

Then Pangram's post does the same thing. It says the origin of an evaluation set and its possible biases can't be seen from outside, and later it reports an internal benchmark of GPTZero run on a subset of Pangram's general set, finding 2.01% rather than the 1% figure Pangram attributes to GPTZero ². One vendor is auditing another on data it doesn't publish.

So there are two unpublished sets and two winners. A school that buys on either table has bought a claim it can't audit, from a company with a stake in the result, about a product version that may already have been replaced. The figures aren't necessarily wrong; they're unverifiable. For a procurement file, that is the worse problem.

Nobody outside the two companies can reconcile the two results. Detector accuracy is two numbers rather than one, and both move with the corpus underneath.

GPTZero's own site makes the point without help from a rival. Its FAQ reports 99.1% of human articles classified as human and 98.4% of AI articles classified as AI, but only inside the high confidence band ⁵. That's a different pair of numbers from the 99.6% and 0.13% on the comparison page. It was measured under a different condition, and both live on gptzero.me.

A peer-reviewed study in the International Journal for Educational Integrity, published in June 2026, compared GPTZero, Pangram, Copyleaks and Turnitin on 160 constructed academic documents with known authorship. The set mixed fully human, fully AI, hybrid and humanized papers. Pangram did consistently better than the other tools on the AI, hybrid and humanized ones, and all four correctly identified the fully human texts ⁸. The authors' conclusion is that detection tools shouldn't be the sole evidence in a high-stakes decision ⁸.

Epoch AI, a research firm in San Francisco ⁴ that sells neither product, published its own test on 15 July 2026. It scored Pangram 3.3.2, a May 2026 GPTZero model and Originality.ai ³. Neither Pangram nor GPTZero produced a single false positive on 495 human-written passages. On AI text written to imitate a named author, Pangram missed 30 of 297 passages, 10.1%, and GPTZero missed 32, 10.8% ³. Thirty against thirty-two isn't a difference a department can act on.

The rest of that result matters more than the ranking. Scientific writing was the hardest genre for all three detectors. Across them, about 26% of the imitated scientific passages went uncaught: Pangram missed 25 of 99 and GPTZero 24 of 99 ³. If graduate coursework reads like the scientific writing Epoch imitated, which is an assumption about the coursework and not a finding of the test, that shared blind spot sits right where that coursework lives.

A 2026 Carnegie Mellon paper found a second one. Text from base language models is often judged overwhelmingly human by GPTZero and Pangram alike, while output from the instruction-tuned versions of the same models isn't ⁷. Neither vendor table predicts either finding.

Nature reached the general version of this in August 2026. Only the accuracy rates the firms announce from internal testing can be current, it said, and those aren't externally verified ⁴. A school reading two vendor pages is reading two unaudited claims, however precise the decimals look.

The limits each company publishes about itself, since nothing cited here shows either side disputing them. Pangram's model card sets a floor of 50 words in complete sentences. It also names the writing the model isn't built for, among them source code, reference sections, templated writing and text dominated by mathematical notation ⁶. GPTZero's FAQ says its results should not be used to punish students ⁵.

Both companies also concede the failure a purchase order usually assumes away. Pangram's model card states that false accusations of AI use can lead to reputational damage, emotional trauma and other undue harm, and that the model has a non-zero error rate ⁶. A competitor's table will never carry that sentence.

One more spread deserves a read before anything is signed, and it comes from a single company. Pangram's false positives post gives roughly 1 in 10,000 overall ². Its model card gives 0.0041%, about 1 in 24,000, on a million human-written FineWeb documents ⁶. Same vendor, two figures, and one is about two and a half times the other. The post's is a rounded overall rate it says moves with the kind of writing ², and the model card's is on a named set. By the vendor's own account, what sits under the detector moves the number.

The same discipline applies to a Pangram and Turnitin comparison, where the two companies also measure on different populations of documents.

Who measured the figure, and on whose data? How many documents, and is the set published? Which model version was tested, and on what date? What does the vendor say the tool can't do? And what happens, procedurally, when a flagged student says the work is theirs? Neither vendor page answers all five, so put the missing ones to the vendor directly.

Ask the fifth one first. It decides whether the other four matter.

A detector that is 99.6% accurate on a vendor's own corpus still produces flagged papers every term. The policy around the flag decides what a wrong reading costs a student, and the decimal in the table doesn't. Pangram's model card concedes the non-zero error rate in writing ⁶, and GPTZero's FAQ concedes the punishment question ⁵.

Human's numbers are a maker's numbers too. Held to the first of those five questions and the size half of the second, Human answers with its own in-house measurement on its own held-out sets rather than an independent audit, measured September 2026: at the standard point, which flags a paper above 6% machine content, 0 wrongly flagged by Human's default model of 600 human-written papers it had never been trained on; the ceiling on that is 0.498%. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. The count describes documents, never students, and nobody outside Human has rerun it.

The benchmark worth trusting is the one a school can rerun.