Super vs Standard vs Lite: Which StealthGPT Humanizer Model Should You Use?

Choose Super when detector-resistant humanization is your priority. In our fresh test of three academic passages, Super cleared the recorded majority-human criterion on all three Pangram 4.1 scans, all three Originality.ai 5.0 scans and two of three GPTZero scans. Standard and Lite did not clear that criterion on these passages. For a faster, lighter edit, Lite had the shortest recorded response time; Standard offered a middle timing-and-credit profile.

We ran the same argumentative essay, laboratory discussion and research-paper passage through all three StealthGPT humanizer models. Nine completed API calls produced nine actual outputs, followed by 36 scans across the three originals and nine rewrites. Below, you can inspect the detector results, matched excerpts, recorded response times and the final edits each example needed.

Super vs Standard vs Lite at a glance

These are observations from three passages per model, all in quality mode. API credit-rate equivalents are the recorded wordsSpent value multiplied by the published base rate of $0.0002 per credit. The calls used prepaid credits; these figures are not separate cash charges or consumer subscription prices. Response time is elapsed time observed by the test client, including network and service processing.

Ready to compare your own draft? Open the StealthGPT AI Humanizer. The AI Humanizer feature guide explains the wider student workflow, including document tools and editing.

How we compared the three models

On November 7, 2026, Pacific time, we sent three original English academic editing fixtures to the live StealthGPT API: a 349-word argumentative essay, a 352-word laboratory discussion and a 357-word research-paper limitations passage. The laboratory and survey data were fictional teaching examples, explicitly identified that way in the source text. These were test passages, not customer submissions.

Every model received the same text through /api/stealthify with rephrase set to true, qualityMode set to quality and outputFormat set to text. We used the first completed result from each model and did not reroll or select outputs by score. The three model requests for each passage ran concurrently. This comparison therefore describes those recorded requests rather than a guaranteed response-time target.

The prose observations below are a Codex-assisted editorial review against the source passages and their predeclared claim checklists. They are not a validated academic-quality score. Internal API detection estimates are excluded from the model-selection table.

Fresh detector results: Super led on all three academic passages

We sent the exact saved originals and outputs to GPTZero, Originality.ai and Pangram. All 36 external scans completed, and text hashes verified that each detector received the same saved text. The original passages cleared none of the three detectors. Super changed that outcome on all three passages for Pangram and Originality.ai, and on the essay and research passage for GPTZero.

These are counts from three fixed passages per model, using the first completed output without rerolls. They describe this demonstration, not an estimated real-world pass rate. The observed detector versions were GPTZero 2026-11-04-base, Originality.ai model-5.0 and Pangram 4.1. Originality’s API served model-5.0 when Turbo was requested, so we report the returned AI-assisted-text measure as Originality.ai 5.0.

Chart calculated from the saved external responses. Download all 36 normalized scan results or the complete input, output and evidence bundle. Each detector’s criterion and observed version are recorded below.

The actual Super scores, passage by passage

Higher is better in the GPTZero and Pangram columns; lower is better in the Originality.ai column. These measures have different meanings and are not averaged together. The laboratory output is the visible exception: it cleared Originality.ai and Pangram but remained below the GPTZero threshold. Turnitin was not part of this test.

For a wider competitor comparison, the student humanizer study analyzes the separately dated October academic-writing subset. This model guide contains the fresh November outputs and scans.

Example 1: an argumentative essay about library hours

The source argues for a limited trial of later library opening: one study floor until midnight on two weekdays for six weeks. Its core distinction is practical access rather than a blanket assumption that all students prefer studying at night.

Original opening

A university library should test later opening hours before making a permanent change. The strongest reason for a trial is not that every student prefers working at night. It is that a single closing time affects students differently.

Lite: a lighter edit

A university library ought to test later opening hours before committing to a permanent change. The best reason for a trial isn't that every student wants to work at night. It's that a single closing time hits students in different ways.

Lite kept the original recommendation at the start, used more conversational phrasing and retained four paragraphs. Its 339-word output was slightly shorter than the 349-word input. The proposed trial conditions and the distinction between observed use and improved grades remained in the text.

Standard: a fuller restatement

A university library should try out later hours for a period before deciding if it is a good idea. One important reason for this trial is not that all students want to work at night. The reason is that a single closing time will affect different students in different ways.

Standard expanded the passage to 390 words while retaining four paragraphs. One edit needed attention: the original said students might face a less reliable journey home; Standard changed this to travelling home in unsafe circumstances. Reliability and safety are different claims. A source check should restore the intended point.

Super: a more substantial reconstruction

There may be good reasons why the library opening hours should be extended and students may well wish this was the case. However, blanket statements about what time the library should close are not helpful.

Super rebuilt the opening and reorganized the argument into a 423-word output. The limited-trial proposal appeared later, after discussion of student needs and operating costs. This is the strongest example here of choosing Super for a more substantial reworking of an argument. Review the new ordering against your assignment’s purpose before using it.

Research and lab reports: check meaning alongside the numbers

All three models retained the laboratory passage’s final mean heights of 11.2 and 13.6 centimetres and their 2.4-centimetre difference. Lite stayed closest to the original sentence progression; Super reconstructed the discussion more substantially. The complete laboratory and research-paper outputs are included in the downloadable evidence pack.

Our editorial review also identified corrections to make before reuse. For example, Super changed the survey outcome from “missed” a class to “late to or missed” a class. Standard changed a recommendation for more than one tray per experimental condition to a separate tray per condition. The per-output review records these details, including fictional-data labels, so you can inspect both the rewrite and its final editing needs.

For research writing, keep each figure attached to its exact population, outcome and timeframe. Use the original method and source material for that final check; the external detector results above measure classification rather than factual accuracy.

Recorded timings, output length and API usage

Output words are counted by splitting the saved output on whitespace. wordsSpent is the API’s billing field, not the output word count. Each response reported prepaid billing and zero meteredChargedCredits. The API reference explains the request fields; see StealthAPI for current API access and pricing.

Which StealthGPT model should you choose?

• Choose Super for detector-focused humanization and more substantial reconstruction. It was the only model in this test to clear any of the three external detectors: all three outputs cleared Pangram and Originality.ai, and two cleared GPTZero. Its outputs were longer and reorganized the argument more extensively. Review the rewrite against the source as part of your final edit.

• Choose Standard for a fuller rewrite with a middle timing-and-credit profile. It completed each request faster than Super and used fewer credits. It did not clear the external detector criteria on these three passages. Check paragraph layout and research terminology when finishing the edit.

• Choose Lite for a quick, lighter revision of an already well-organized draft. It had the shortest observed response time and lowest recorded credit use on all three passages. It stayed close to the source in the essay and research examples, but did not clear the external detector criteria here.

A useful approach for an essay is to humanize one section, compare it with your source and assignment requirements, then use the model whose changes best fit your editing goal. The research-paper guide walks through a source-led revision process. For automated document workflows, see our API document-preservation test.

Frequently asked questions

Which model should I choose for AI detection concerns?

Start with Super. In this matched three-passage test, Super cleared Pangram and Originality.ai on all three outputs and GPTZero on two. Standard and Lite cleared none of those detector criteria. Keep the actual output and source comparison as part of the final edit.

Which model was fastest in these tests?

Lite was fastest on each of the three passages. Its median recorded response time was 2.185 seconds, compared with 7.687 seconds for Standard and 18.861 seconds for Super. These are recorded API timings for these requests.

Are these the prices students pay in the web app?

No. The dollar figures are API credit-rate equivalents calculated from recorded API usage. They do not describe the student subscription price or a separate cash charge on the prepaid test account. Check current plans for web-app access.

Where can I inspect the original texts and complete outputs?

Download the complete model evidence pack for all three original passages, all nine unedited outputs, external detector evidence, settings, usage and per-output editorial review. The separately dated academic humanizer study compares 11 services using October benchmark data.

Try the model that fits your next revision in the StealthGPT AI Humanizer, then check the finished passage against your sources and the point you intend to make.