Pangram AI Detector: Free Limits, Pricing & Review
Choose Pangram when you want passage-level analysis of longer prose and a documented daily free allowance. Test your genre first; short answers, reference lists and code are a weak reason to buy it.
Source checked October 19, 2026. Editorial research; no original accuracy benchmark or hands-on score is claimed.
What Pangram does
Pangram is worth evaluating for editors and educators who need to inspect passages rather than accept a single document score. Its Pangram 4 model card describes human, AI-assisted and AI-generated segments. The model targets long-form natural language with at least 50 words; code, templates, references and very short answers are poor matches. A label is an estimate of writing origin, not proof of misconduct.
Free access and paid plans
The official free-account guide lists 20 daily credits, equivalent to up to 2,000 text words, plus three image checks per day, without a payment method. Its help article was last reviewed July 24; confirm the allowance in your account. The current pricing page advertises Individual at $20/month after a seven-day trial, with 300,000 words and 100 images monthly. A subscription trial is separate from the limited free account.
Compare before choosing
Free AI detectors: limits, privacy and false positives · GPTZero vs ZeroGPT · All AI content detectors
Read the result at the right level
In a document assembled from a human draft, generated suggestions and manual revision, a single whole-document label hides the question that matters: which passage warrants review, and why? Pangram’s segmented result can help organize that review. It cannot show an author’s intent or whether a particular type of AI assistance was permitted. Set those rules before checking a submission.
For example, an editor reviewing an article can compare a highlighted paragraph with interview notes, citations and the earlier draft. If the paragraph is a standard methodology description, formulaic wording may be expected. If it introduces a quotation absent from the recording, the missing source is a concrete editorial issue irrespective of the detector label.
Privacy: storage, training and repositories are separate
Pangram’s training-data help says submissions are not used to train its models, but content and results are stored for later viewing. Its repository article says uploads are not added to a shared plagiarism database. Its zero-retention article places that option in custom enterprise arrangements. These are vendor statements, not a security audit. “Not used for training” should not be read as “not stored.”
A practical Pangram pilot
There is positive historical evidence worth reading: the University of Chicago’s October 2025 research brief reports strong Pangram results in a four-detector comparison and emphasizes threshold tradeoffs. It is not a test of every present-day language, genre or release. The Pangram 4 technical report (2026) is a separate model report; evaluate its benchmarks as evidence about its disclosed test conditions, not a guarantee for your next document.
- Pick the job. Use editorial prose or essays similar in length and subject to the documents you actually handle. Exclude code and lists from your main trial.
- Preserve provenance. Keep human draft history and record the model, prompt and edit history of generated examples. Do not decide the answer from another detector’s score.
- Inspect extraction. Check the pasted or uploaded text against the original. Broken PDF columns and missing paragraphs change the input.
- Record passage-level utility. Note whether the result points to a reviewable passage, not just whether it returns a confident label.
- Price the complete workload. Include repeat scans and image checks in the appropriate quota. Do not treat a daily word allowance as a monthly pool that can all be spent on one day.
When to compare an alternative
| Requirement | What to evaluate next |
|---|---|
| Draft history is central | GPTZero and its writing-process features; confirm your workflow is supported. |
| AI plus matching existing sources | Copyleaks; compare report quality and actual credit consumption. |
| A quick public text check | ZeroGPT; input capacity is a separate question from reliability. |
The lowest subscription price is not necessarily the lowest review cost. A service that flags many legitimate documents may require more staff time than the scan fee suggests. Track minutes spent resolving each flag during your pilot.
Download our blank detector evaluation worksheet (CSV) and purchase checklist. They contain no detector results. Fill them with your own authorized samples.
Frequently asked questions
Is Pangram free or just a free trial?
Its help center documents a limited free account, while paid Individual access has a separate trial. Read the free-account allowance and the subscription renewal terms as different offers. Signing up for a trial does not establish that it will stay free.
Does “AI-assisted” mean the author cheated?
No. The classification describes the detector’s estimate, while the relevant rule concerns what assistance was allowed and disclosed. Review the passage, ask for drafts and apply the published policy. The same editing assistance can be acceptable in one workflow and prohibited in another.
Can a no-training promise cover confidential client work?
Not by itself. Confirm retention, authorized access, the contract you need and the deletion process. Pangram documents a separate custom zero-retention option; verify that it is enabled for your agreement rather than assuming it applies to a free account.
Is Pangram more accurate than GPTZero?
We have not run a matched current-version benchmark, so this page does not award an accuracy winner. Use a labeled set from your own language and domain and report false positives, misses and uncertain cases separately.