Vendor benchmarks: what GPTZero and Pangram publish about their own accuracy

The record joins the studies index beside the Turnitin claims record as a vendor-statement record: it documents what two companies publish, dated and versioned, so the claim vs evidence gap ranking can set each claim against the studies that tested the same tools. Vendor figures are never averaged with study figures in this index.

GPTZero scores itself and four competitors; Pangram scores itself and, in its 2024 technical report, two competitors. Every detector named has a card (GPTZero benchmarking, Table 1; arXiv 2402.14873, Section 3.1, read 9 November 2026).

Versions as the vendors print them. GPTZero says Turnitin could not be scored because access is limited to educators and that "Copyleaks has blocked our account" (GPTZero benchmarking, Evaluation Accessibility, read 9 November 2026).

GPTZero selects human texts from its own database and generates an AI version of each; each English benchmark holds 1,000 human and 1,000 AI texts, the AI texts from GPT-5.2, Gemini 3 Pro, Claude Sonnet 4.5 and Grok 4 Fast (GPTZero benchmarking, Benchmark Construction, read 9 November 2026). Pangram measures false positives on pre-ChatGPT FineWeb web text, cleaned for 4.1 to remove "spam, advertising and programmatically generated documents" (Pangram 4.1 model card, Evaluation sets, read 9 November 2026).

Counts as printed by each vendor on 9 November 2026.

Both vendors count a Mixed prediction as AI when scoring (GPTZero benchmarking, Metrics and Binarization; Pangram 4.1 model card, Methodology, read 9 November 2026).

Every figure is a vendor figure, printed on the page or card named in its row (GPTZero benchmarking; Pangram 4.1 model card; Pangram 4 model card; arXiv 2402.14873, read 9 November 2026).

All values as printed by the vendor named in each row (Pangram 4 model card, read 9 November 2026).

GPTZero's 0.08% is an average over four domains of 1,000 human texts each, so a per-domain 0.00% means no false positive in 1,000 (GPTZero benchmarking, Tables 2 to 5, read 9 November 2026). Pangram's 0.0090% rests on 781,981 human texts, and the card says the headline rate is calibrated "to roughly 1 in 10,000" by design, so the number is an operating point the vendor set (Pangram 4.1 model card, Evaluation sets, read 9 November 2026).

Each vendor also scores the other. GPTZero's table places Pangram 3.2 at 95.66% recall against its own 99.60%, and at 68.11% on bypassed text against its own 91.80% (Tables 1 and 7). Pangram's 2024 technical report placed GPTZero at a 10.02% false negative rate and Originality at a 9.24% false positive rate on 1,976 documents, claimed "over 38 times lower error rates" for Pangram, and reported that GPTZero's October 2023 model flagged 7.7% of the 91 TOEFL essays from Liang 2023, "or 1.1% if 'Possible AI Content' is generously labeled a negative" (arXiv 2402.14873, Sections 3.2 and 3.5, read 9 November 2026). GPTZero's technology page gives the 1.1% TOEFL figure as its own (GPTZero technology page, read 9 November 2026).

Pangram's 4.1 card also lists reruns of its current model on third-party data, among them 0.000% false positives and false negatives on the 15,936 full-length texts of Jabarian and Imas 2025 (Pangram 4.1 model card, 3rd party benchmarks, read 9 November 2026). Those are vendor reruns, not the published study results. Two question pages added 9 November 2026 read this record: the repeatability question, for the rerun and the model versions, and the published test data question, for which vendors name or size their test sets.

GPTZero states that three of its four English datasets cannot be published because of "data usage policies and copyright constraints" and that detectors whose APIs time out have those texts excluded from their metrics (GPTZero benchmarking, read 9 November 2026). Pangram states that its model "has a non-zero error rate" and that code, markup, reference lists, templated writing and AI translations are out of scope (Pangram 4.1 model card, Intended usage and limitations, read 9 November 2026).

Evident from the sources: each vendor chose the test texts, the generators, the competitor versions and the binarisation rule, and each ranks itself first; GPTZero's per-domain false positive rates move in steps of 0.10%; Pangram's human base is pre-2022 web text, not student writing; and Pangram's headline rate changed from 41 in 1,000,000 to 70 in 781,981 between July and October 2026 because the test set changed, not only the model.

Both vendors' dated model pages also appear on the model changelogs view, added 10 November 2026, beside the eight other cards whose vendor publishes dated release notes, a model card or a version post.

Two cards take a dated vendor figure and three take a note that a competitor scored them (GPTZero benchmarking; Pangram 4.1 model card, read 9 November 2026).

None of these figures enters the false positive rate ranking or the false negative rate ranking, which take study results only; the claim vs evidence gap ranking is where vendor claim and study result meet. How this record is sourced explains the rule; the home page shows which AI detectors exist and who has tested them.

Related records: Epoch AI 2026 scored Pangram 3.3.2 and GPTZero 2026-05-11-base at 0 of 495 false positives each on an outside sample; DAMAGE 2025 is Pangram's own research paper on humanizers.

Both vendor documents in this record qualify as a model card under the model card question, added 9 November 2026: a document naming the model version or date, the test set and its size, and the metric. That page finds 8 such cards among the 72.