Pangram Lists No School Price; an Institutional License Is a Quote

Nothing that Pangram Labs publishes. The pricing page sells individuals and small teams by the word, prices an API for developers, then stops. The Educational Institution row names schools, districts and universities, describes an institutional license, and puts a contact form where the individual, team and API tiers show a figure ¹. Any school price is settled in a conversation rather than on a page.

The listed tiers, read on 18 September 2026 ¹:

Annual billing is listed as a saving on the individual, professional and team rows, and bulk API use carries a 20 percent discount ¹. Two rows carry no figure at all.

Eight seats on the Team plan would list at $160 a month ¹. A price under a card limit does not settle whether a department may send student papers to a vendor; that is the school's procurement and records decision.

Stare at the API row. Pangram 4 bills ten times as much per word as Pangram 3, and both models are on sale at once ¹. Ask which model the license runs, and what happens to the price when the model changes.

Pangram's free tier runs 2,000 words a day ¹, or three essays of about 650 words, so it prices a trial rather than a department.

Five feature lines and a contact form, and no number. That is the whole of the Educational Institution row for schools, districts and universities on the pricing page ¹. Its universities page adds that Pangram Labs is FERPA and SOC 2 Type 2 compliant and trains on neither student papers nor proprietary research ².

The five, read on 18 September 2026 ¹. First, "Unlimited AI checks and plagiarism checks integrated directly into the LMS". Then automatic plagiarism detection, covering copy-paste work and re-used submissions from past years. Then visibility into usage, trends and statistics, and full regulatory compliance and data controls with no training on student data. Last, integrations in Canvas, Brightspace, Moodle, Google Classroom, Google Docs and more.

What the page leaves out is the part a budget needs: the unit the price counts and the contract term. The integrations are named. Whether every named integration is included at the quoted figure is not. Plagiarism detection is listed as part of the license ¹, which is a line on a marketing page rather than a line in a contract.

Hidden is the wrong word for all of this. It is unpublished, which means the first real number arrives after a sales call.

Pangram inside Canvas SpeedGrader is one of the named integrations, and "integrated directly into the LMS" covers a range of work that a quote has to itemize.

The compliance line and the license are one conversation, not two. Under FERPA, the United States federal student-records law, a school may treat an outside vendor as a school official on three conditions ⁴. The vendor performs a service the school would otherwise staff itself. It stays under the school's direct control over the use and maintenance of education records, and it is subject to the limits of 34 CFR 99.33(a) on use and redisclosure. Guidance from the U.S. Department of Education, published in February 2014, lists a further element: the provider has been determined to meet the criteria for a school official with a legitimate educational interest set out in the school's annual FERPA notification ⁵.

A notice that already states criteria for a school official can cover a vendor that meets them; the determination is the step, not fresh wording ⁵. How the rule applies to one district is a question for the school's own counsel, not for a vendor page or an article.

What FERPA asks of a detector vendor is a legal question no pricing page answers.

Because a figure is a reading on a date. Pangram Labs released Pangram 4 on 29 July 2026 ³. The pricing page read on 18 September 2026 bills that model at $0.05 per 100 words, against $0.05 per 1,000 words for Pangram 3, with both on sale at once ¹. A summary written before that release can be accurate on its publication date and out of date as a budget input later.

So a figure lifted from a comparison site or a conference slide is not the vendor's price. It is what the vendor's price was on the day that page was written, and a page without a read date cannot be checked.

One column in the spreadsheet fixes this: write the date the figure was read beside the figure. Every number on this page carries 18 September 2026 for that reason ¹.

Re-read the vendor page before a number reaches a board paper, and read it again before a renewal.

What colleges are actually using turns over on the same cycle, and for the same reason.

Count the submissions first, because unlimited describes checks, not consequences. A district sending 75,000 papers through a detector is buying 75,000 readings, and the number of wrong flags scales with the human-written papers in that count. Vanderbilt University did that sum in 2023 on the 75,000 papers it sent to Turnitin in 2022: at the 1 percent false positive rate it attributed to Turnitin, around 750 papers could have been incorrectly labeled ⁶.

That sum treats all 75,000 as human-written and takes the vendor-stated rate as holding on Vanderbilt's papers; both are assumptions.

Count per term, before the call, and carry the shortest assignment in the catalog along with the total.

Length is the other half of the count. Pangram Labs states that Pangram 4 is built for prose of at least 50 words in complete sentences. Source code, reference sections, tables of contents and templated writing sit outside its scope ³. Neither the pricing page nor the model card says whether an LMS integration filters or labels submissions outside that scope ¹³. It is a quote-call question.

The same sum can be run on any vendor's published figure, Human's included. Human's own reading, as of September 2026: 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%. That is an in-house measurement on Human's own held-out set, not an independent audit, and a district should weigh it the way it weighs any vendor's. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. Applying a rate measured on one set of documents to a district's own papers is an assumption, and the sum belongs on the human-written papers, not on every submission. The measurements sit at https://human.olive.is/research/model-card, which names the run behind them.

The false positive arithmetic is the same sum for any vendor's figure, and it is worth doing on paper before a license is signed.

Start with the unit, because everything else prices off it. Whether the license counts enrolled students, staff seats, submissions or courses decides what a renewal looks like after a strong enrollment year, and the pricing page answers none of it ¹. Ask for the term and the renewal cap in the same message. Ask what an exit costs.

The list worth sending ahead of the call:

• The unit and the count: enrolled students, staff seats, submissions or courses, and who audits the number.
• Term length, the cap on a renewal increase, and what a mid-term enrollment change does to the invoice.
• The page names integrations in Canvas, Brightspace, Moodle, Google Classroom and Google Docs ¹; which of those are included at the quoted figure, and which arrive as a separate line.
• Written confirmation that plagiarism checking, which the row lists as part of the license ¹, sits inside the quoted figure.
• Whether the LMS integration filters or labels submissions outside the model's stated scope, such as source code and reference sections ³.
• Retention: how long submitted papers are stored, and what a district does to delete them.
• How the vendor expects to be treated under FERPA's school-official exception, in writing, and whether it meets the criteria in the school's existing annual notification ⁵.

An accuracy figure offered on that call may be the vendor's own internal measurement. Nature's assessment of this market is blunt about the limit: only the rates firms announce from internal testing can be truly current, and those are not externally verified ⁷. Ask for the evaluation set rather than the headline, and for its size and its date.

Pangram Labs is plain about the stakes. Its model card states that the model has a non-zero error rate, and that false accusations of AI use cause serious harm including reputational damage and emotional trauma ³.

Read that sentence aloud in the meeting where the license gets signed.

The full vendor question list runs past price, and price is the easiest part of it to get in writing.