GPTHuman Review: Pricing, Limits, and a Preservation Test
GPTHuman gives buyers enough public detail to price a job, but its pages use different language counts, trial descriptions, and Plus allowances. Its own Help Center also says scores cannot verify facts or citations. This review turns that warning into a fixed quotation, number, and table fixture that a reader can rerun.
HumanPen Team
· 13 min read
Is GPTHuman worth trying?
A short, non-sensitive GPTHuman trial is reasonable when copied text is an acceptable output. Its Free plan advertises 300 words per output, while paid plans raise both the monthly output pool and the per-output ceiling. The service was not run for this review, so no detector result, preservation result, or rewriting-quality claim is established here.
The useful starting point is GPTHuman's own Help Center answer to Do AI Humanizers Actually Work?. It keeps readability, similarity, factual accuracy, and detector results apart because they measure different things, instead of folding them into one success score. Asked whether a humanizer can guarantee a detector result, the page answers "No. AI detectors use different datasets, models, thresholds, and definitions." The same page counts changed meaning, an altered name, number, date, quotation, or citation, and unsupported facts among the ways a rewrite can fail.
Short answer: use the free allowance for a fixed evidence test, not as proof of a future detector outcome.
What does GPTHuman itself say the scores cannot prove?
The output-quality article describes Human, Similarity, and Readability scores as three separate signals. It says none independently verifies factual accuracy, quotations, citations, plagiarism, policy compliance, authorship, or how an external detector will classify the text.
| Public guidance | Boundary it publishes | What this review does with it |
|---|---|---|
| Meaning preservation | Central meaning is the design goal, but exact preservation cannot be assumed in every passage | Gives exact facts, conditions, a quotation, a citation, and arithmetic to compare |
| Output quality | Similarity and readability are review signals, not approval decisions | Scores stay outside the pass criteria |
| Do AI Humanizers Actually Work? | No humanizer can guarantee a detector result, and the Human Score is not proof of human authorship | Reports no detection rate and runs no detector |
| Responsible use | Users should verify facts, calculations, quotations, and citations and follow the applicable rules | Uses those four items as the acceptance fixture |
This review uses a narrow, testable boundary: a vendor score cannot approve evidence for you, and no public result claim is treated as an independently reproduced result.
A fluent paragraph with a high internal score can still contain one wrong number or one citation attached to the wrong claim.
What goes in and what comes back?
The current homepage DOM exposes one textarea and zero file inputs. The account workflow is broader according to How to humanize content using GPTHuman: users can paste text or import `.docx` and `.pdf`, after which the service extracts the text.
| Stage | Publicly documented route | Boundary to keep |
|---|---|---|
| Homepage sample | Paste text into a browser textarea | A public sample surface, not the paid account ceiling |
| Account input | Paste text or import DOCX/PDF for text extraction | File acceptance does not by itself establish a same-file return |
| Processing | Select a tone and mode; the account counts words | No setting was exercised in this review |
| Result | Text appears in the right panel and can be copied; re-humanizing is a separate action | The checked guide does not describe downloading the humanized DOCX or PDF |
The help article names Professional, Balanced, and Enhanced modes. Its meaning-preservation guidance says a conservative rewrite generally stays closer and that stronger variation creates more opportunity for drift. Those are vendor guidelines, not observed results.
Follow both ends of the route: importing a DOCX can still lead to copied text rather than an editable DOCX round trip.
What fixture can you copy and run?
Use the same non-sensitive source every time. The paragraph below contains 83 whitespace-delimited tokens when every run of non-space characters counts as one token. All names, events, and figures are invented for this test.
Rivera (2025, p. 12) wrote, "Batch A stayed below 4.75 ms." The audit covered 128 records on 12 March 2025. Batch A contained 64 records and 9 exceptions; Batch B contained 64 records and 8 exceptions. The combined exception rate was 13.28% (17/128). The rate applies only when room temperature is at or below 21.5 C. Because the sensor failed after record 96, records 97-128 were excluded from the latency average but remained in the exception count. The conclusion is preliminary, not final.
Paste this table immediately after the paragraph:
| Cohort | Records | Exceptions | Rate |
|---|---|---|---|
| Batch A | 64 | 9 | 14.06% |
| Batch B | 64 | 8 | 12.50% |
| All | 128 | 17 | 13.28% |
The fixture is intentionally much shorter than 300 words, but GPTHuman defines the Free limit as words per output, not words in the source. A short input leaves room; it does not guarantee how long a generated result will be.
How should you score the fixture?
Build the baseline before submitting anything. Pass or fail each checkpoint separately rather than averaging them into one quality score.
| Checkpoint | Baseline | Pass condition |
|---|---|---|
| Citation and quotation | `Rivera (2025, p. 12)` immediately introduces one quoted sentence | Attribution, locator, quotation marks, wording, and attachment are unchanged |
| Numbers and dates | Every value, date, range, decimal, unit, fraction, and percent sign in the paragraph and table | No token changes value or role |
| Arithmetic | 9/64 = 14.06%, 8/64 = 12.50%, and 17/128 = 13.28%, rounded to two decimal places | Recalculation still matches every displayed rate |
| Condition | The rate applies only at or below 21.5 C | `only` and the threshold direction survive |
| Exclusion rule | Records 97-128 leave one average but remain in the exception count | The two populations are not merged or reversed |
| Qualification | The conclusion is preliminary, not final | Uncertainty is not upgraded into certainty |
| Table shape | Four headers, three data rows, twelve body cells, fixed row order | Shape, labels, order, and every cell value survive copy and paste |
- Save the source as plain text and keep an untouched copy.
- Record the language, tone, mode, plan, time, and whether the source was pasted or imported.
- Run only one setting at a time and copy the result into a new plain-text file.
- Check exact tokens first, recalculate the three rates, then compare the condition, exclusion rule, and qualification.
- Record each failure by checkpoint. Do not replace the ledger with Human, Similarity, Readability, or detector scores.
This protocol measures evidence retention in one authorized sample. It predicts no detector result.
What do the plans actually cost and allow?
The USD pricing page was checked in both billing states on August 28, 2026. Its FAQ defines `words per month` as Humanizer output generated during a billing cycle and `words per output` as the maximum length of one result.
| Plan | Month-to-month view | Yearly view: annual charge and displayed monthly average | Output pool | Maximum per output |
|---|---|---|---|---|
| Free | Free trial | Not applicable | Trial access: 300 words | 300 words |
| Starter | $15/month | $99/year; page displays $8.25/month | 25,000 words/month | 1,000 words |
| Plus | $25/month | $168/year; page displays $14/month | 60,000 words/month | 1,500 words |
| Unlimited | $49/month | $312/year; page displays $26/month | Unlimited words/month, subject to abuse guardrails | 2,500 words |
The yearly monthly figures are display equivalents attached to annual charges. They are not the month-to-month prices. `Unlimited` also does not remove the 2,500-word per-output ceiling or the published abuse guardrails.
Keep billing cadence, full charge, monthly output pool, and per-output ceiling in separate columns.
Which public statements need reconciliation?
Several current first-party pages describe the same purchase with different labels or figures. Preserve each page's wording rather than choosing the largest number.
| Question | First public statement | Second public statement | Safe reading |
|---|---|---|---|
| Languages | Homepage says 80 languages | Pricing and API pages say 50+ | Test the exact language; no uniform count is established across pages |
| Plus allowance | Pricing says 60,000 output words/month | Trial help article says 100,000 credits/month | Credits and output words are not interchangeable without a current account rule |
| Starting paid price | Pricing help article says subscriptions start at $8.25/month | Pricing page attaches $8.25/month to $99/year; the Monthly view says $15/month | `$8.25` is the yearly plan's displayed monthly average on the checked page |
| Trial | Pricing says 300 words per output; trial help says 300 credits with no card | Terms contain a seven-day free-trial clause | Confirm whether the current account uses a word balance, credit balance, time window, or a combination |
The account guide says credit cost is typically one word per credit, but it also separates Unlimited humanization from credits. `Typically` is not enough to convert every displayed credit figure into an output-word entitlement.
When two official pages differ, keep both page names, dates where available, units, and billing states in the note you take to support.
What do cancellation and refund terms say?
The Terms and Conditions are marked last updated December 1, 2024. Their fees section authorizes recurring charges until cancellation. Under the cancellation section, purchases are non-refundable and the paid term must expire before cancellation takes effect.
| Contract question | Public wording checked |
|---|---|
| Renewal | Recurring charges continue until the user notifies GPTHuman of cancellation |
| Refund | Terms say all purchases are non-refundable |
| Cancellation route | Account or the contact details in the Terms |
| Effective date | When the current paid term expires |
| Help Center steps | Billing, Next invoice, Cancel subscription, then follow the prompts |
| Remaining access | Cancellation help says features and credits remain usable until the next billing date, then expire |
This means `cancel anytime` describes when a request can be made, not an immediate refund. The Terms' seven-day trial clause and the current no-card 300-credit trial description should be confirmed in the account before relying on either one. This article reports the written pages and offers no legal advice.
Check the payable total and cancellation date before buying an annual plan; the purchase is described as non-refundable.
What do the privacy pages say about submitted content?
The Privacy Policy, last updated April 14, 2025, says User Content includes prompts and uploaded files. It says content may be used to improve the service and train models unless the user opts out in account settings, and it gives no single fixed retention period.
| Surface | Statement checked | Scope boundary |
|---|---|---|
| General Privacy Policy | Collects prompts and uploaded files; may use content for model training unless opted out | Applies broadly to the Services described by the policy |
| Homepage API block | Says API content is never stored or reused | Appears inside the Developer API section |
| API product page | Says API content is not shared, stored, or used for retraining | API-specific product statement |
| Retention section | Keeps personal data as long as needed for service or other legitimate purposes | No universal deletion interval is printed |
The API-specific no-storage language and the general policy do not answer the same scoped question. Before sending confidential, unpublished, regulated, or institution-restricted text, identify the rule for the particular browser or API route and retain the answer.
A privacy promise for the API should not be silently transferred to the browser Humanizer.
Who is this workflow a fit for?
GPTHuman's public contract is clearest for text that can be pasted or extracted from a DOCX/PDF, reviewed as copied output, and kept within a known per-output limit. The decision becomes less clear when the returned file itself, a disputed page allowance, or a fixed deletion deadline matters.
| Need | Verify before relying on it |
|---|---|
| A short, non-sensitive passage | Run the fixed fixture and keep the checkpoint ledger |
| A long job | Per-output ceiling, monthly output pool, billing state, and current account counter |
| DOCX or PDF source | Import behavior, extraction order, tables, citations, and the documented copied-text result |
| Multilingual writing | Exact language in the selected mode; public pages say both 80 and 50+ |
| Evidence-heavy writing | Every number, quotation, citation, condition, exclusion, and table cell |
| Sensitive material | Web-versus-API privacy scope, opt-out state, retention, deletion, and permission to upload |
| Paid use | Monthly or yearly charge, applicable allowance, non-refundable term, and cancellation date |
| Academic work | Institution or instructor rules, required disclosure, and the user's own authorship and review |
One passing fixture establishes one language, mode, input route, and point in time. Store those fields with the output. It does not establish every plan, file type, language, or future detector version.
Choose the workflow whose output you can inspect and whose contract matches the real deliverable.
When is same-format document return more relevant?
HumanPen is both this site's publisher and the vendor behind the alternative below. It therefore has a commercial interest in the comparison. Its public page says each DOCX, PPTX, or TXT comes back as the same file type; Word and PowerPoint outputs stay editable; and the user confirms which passages may change.
| Decision point | GPTHuman public guidance | Publisher's public page |
|---|---|---|
| Input | Paste text or import DOCX/PDF for extraction | Upload an editable document or text file |
| Output | Result appears in a browser panel and is copied | Same-format return; Word and PowerPoint remain editable |
| Rewrite scope | Whole submitted text under the selected mode | User-confirmed passages or passages located from an imported report |
| Evidence here | No run or detector test | First-party disclosure, not an independent output test |
This is a structural comparison, not an effectiveness comparison. Run the same quotation, number, condition, and table fixture on either route, then inspect the final artifact required by the job.
A document workflow matters when editability, layout, and untouched surrounding content are part of the acceptance criteria.
What is the verdict on GPTHuman?
GPTHuman publishes useful operating limits: the Free 300-word output, three paid output pools, four per-output ceilings, separate monthly and yearly prices, and an abuse qualifier on Unlimited. Its Help Center also publishes direct limitations on internal scores and meaning preservation.
The pages are less settled on language count, the Plus allowance, how the free trial is measured, and whether the lowest displayed price is monthly billing or an annual-plan average. Check those differences in the live account or with support instead of guessing.
The fixed fixture is the decision gate. If its citation, quotation, numbers, arithmetic, condition, exclusion rule, qualification, and table all survive, you have evidence for that exact setting and input route. If one fails, the ledger tells you what changed without borrowing certainty from a score.
Price the output, reconcile the contract, and test the evidence.
Frequently asked questions
Is GPTHuman free? The pricing page shows Free trial access with 300 words per output. The trial help article says no card is required and calls the allowance 300 credits. Confirm the live counter before assuming credits and words are identical for every feature.
How much does GPTHuman cost month to month? The checked Monthly view showed Starter at $15, Plus at $25, and Unlimited at $49 per month. The lower $8.25, $14, and $26 monthly figures belong to the Yearly view and sit beside annual charges of $99, $168, and $312.
How many words can GPTHuman process at once? The pricing FAQ defines per-output maxima of 300 words on Free, 1,000 on Starter, 1,500 on Plus, and 2,500 on Unlimited. A monthly pool is not a single-request limit.
Does GPTHuman preserve citations, numbers, and meaning? Its Help Center says central meaning is the goal but exact preservation cannot be assumed in every passage. It tells users to verify names, dates, numbers, quotations, citations, qualifications, and technical terms. Use the fixture above rather than treating Similarity as a guarantee.
Does GPTHuman return an editable DOCX? The account guide describes importing DOCX or PDF, extracting the text, showing the result in the right panel, and copying it. The checked guide does not document a humanized DOCX download. Verify the actual return route with a disposable file if that is required.
Does this review show that GPTHuman passes Turnitin or another detector? No. Neither GPTHuman nor any detector was run. GPTHuman's own testing article says its Help Center does not present a universal independently verified bypass rate and that detector results require a named test set, versions, date, thresholds, and complete outcomes.
Official pages checked
The following first-party pages were checked on August 28, 2026:
- GPTHuman homepage
- Pricing
- Terms and Conditions
- Privacy Policy
- Academic Integrity and Responsible Use
- AI Humanizer API
- What is the pricing?
- How to cancel
- How the trial works
- How to humanize content
- Meaning preservation
- Output quality
- Do AI Humanizers Actually Work?
- Publisher's document workflow
This review created no account and submitted neither text nor files. It did not run a humanizer, score, detector, API, checkout, cancellation, or refund flow. Prices, allowances, page wording, and policies can change; repeat the check before paying or uploading sensitive material.
Source review date: August 28, 2026. Page claims remain attributed to their publisher.
KEEP READING