StealthGPT

What StealthGPT is

StealthGPT is a humanizer that does not pretend to be anything else. You paste text an AI wrote, pick a model and a tone, and it returns a rewrite intended to read as human-written to Turnitin, GPTZero, Originality.ai, Copyleaks and the rest. The homepage headline is "The AI Humanizer that actually works" and a line under the plan cards reads "We beat every AI Detector". There is no softening layer, no ethics page, and no attempt to frame the product as anything but detector evasion.

Three models sit behind it. Super is the flagship, described on the API page as "Built to pass Pangram V4 and major AI Detectors"; Standard is the previous generation, "Proven against GPTZero, Winston and Originality"; Lite is the fast one, "For small human touchups when detection isn't a concern". Tone is a separate axis with four settings, Standard, HighSchool, College and PhD, alongside a quality mode with built-in checks against a fast single-pass mode and a detail control running High, Medium and Low. Every API response carries a detection score from 0 to 100 and the words it spent, which is genuinely useful design: the tool grades its own output before you use it.

The meter is the unusual part. Most of this category sells monthly word allowances; StealthGPT sells weekly, monthly and yearly, and the weekly option is not a rounding of the monthly one. Pro is 12,500 words a week or 50,000 a month, Max is 50,000 a week or 200,000 a month, and a per-request cap sits on top, 1,000 words for Pro and 2,000 for Max, so a long document is several requests no matter how much allowance you hold. Around that core sits a wide accessory surface: a Chrome extension, an iOS app, Stealth Writer and Stealth Agent, an SEO writer, Chat with PDF, Photo to Answers, a grammar checker, simplifier and a watermark remover with a Claude-specific variant, plus three REST endpoints, a public MCP server and a Zapier connector.

What nobody can verify, us included, is the part you are actually buying. Evasion is a moving target: a success rate against any given detector is a snapshot that expires the next time that vendor ships a model. No published figure settles it, and a single-date result would not either. The vendor's own numbers do not agree with each other. The detector and humanizer pages both claim a 99% bypass rate across every major detector; the API page advertises the flagship model at 89% against Pangram V4. Same site, same day, and neither figure carries a methodology, a sample size or a date.

One thing this vendor does that its rivals do not is state the conflict of interest out loud. StealthGPT also sells an AI detector, and its page opens by saying so: "Most AI detectors are built by companies trying to catch AI content. StealthGPT's detector is built by the company that bypasses those detectors for a living." A candid sentence, and also a reason never to treat that tool's score as independent confirmation the rewrite worked.

Pricing

StealthGPT publishes every consumer price without a login, on three billing toggles, and backs it with a support-center article titled "Current plans and pricing" that restates the same figures as standard prices "in USD". That is the cleanest pricing source in this cluster, and the one we used to settle the annual totals. Entry is $6.99 billed weekly for Pro, or $29.99 billed monthly after a first month advertised at $1.00. The terms confirm the currency: "All payments shall be in U.S. dollars."

All figures in US dollars, read from stealthgpt.ai/pricing on all three billing toggles, the support center's "Current plans and pricing" guide (updated 29 August 2026) and stealthgpt.ai/stealthapi on 24 September 2026. Payments run through Stripe. The terms of service state that sales tax will be added "as deemed required by us".

How to read the $1 month, the annual toggle and the word meter

The $1.00 is the first month, not the price. On monthly billing the Pro and Max cards lead with "$1.00*" above "Introductory price · Then $29.99" or "Then $49.99", and the asterisk resolves to "Limited trial until first renewal". This is a paid first month at a dollar, on a card, that renews at thirty or fifty. The weekly toggle carries no introductory offer, the yearly toggle carries a different one, "$5.99 off your first year" on Pro and "$18.99 off your first year" on Max, and none of the three appears in the support center's price list.

The yearly cards and the support guide do not multiply out to the same total. The support center gives standard yearly prices of $299.99, $499.99 and $1,799.99, with monthly equivalents of $25.00, $41.67 and $150.00. The cards print $24.99, $41.99 and $149.99 a month. Pro and Enterprise are a cent of rounding; Max is not. $41.99 twelve times is $503.88, which is $3.89 more than the $499.99 the vendor's own guide calls the standard yearly price.

"Save up to 52.5% with yearly" is a weekly comparison, not a monthly one. The banner sits directly above the billing toggle, but monthly against yearly is about 17 percent on all three plans: $359.88 against $299.99, $599.88 against $499.99, $2,159.88 against $1,799.99. The only pairing that gets near 52 percent is the Max weekly rate against its yearly price, $1,039.48 over 52 weeks against $499.99. That arithmetic is ours; the page does not say which comparison it means.

Budget in words per request, not just words per period. The per-request cap is the limit most buyers hit first: 1,000 words in a single call on Pro, 2,000 on Max, 10,000 or more on Enterprise. A 4,000-word essay on Pro is four requests you then stitch back together. Nothing published says whether re-running the same paragraph deducts the words twice, and on a tool people re-run to chase a score that is the number that decides your real bill. The API is the one place the metering is stated cleanly: words never expire, and every response returns the words spent and the balance left.

The API is the cheapest way to buy words, by the vendor's own numbers. Dividing each plan's price by its allowance gives roughly $0.60 per 1,000 words on Pro monthly, $0.56 Pro weekly, $0.50 Pro yearly, $0.40 Max weekly, $0.25 Max monthly, about $0.21 Max yearly and $0.18 on Enterprise. The API's Standard model is $0.20 per 1,000 words with no subscription and no expiry, which undercuts every consumer plan except Enterprise. In the other direction, Super is quoted at $0.05 per 100 words, which reads cheaper than $0.20 but is $0.50 per 1,000, two and a half times the Standard rate. That arithmetic is ours.

What we verified, and what we did not

Reviews of tools like this often imply the score covers whether the thing works. It does not, so here is the boundary in plain terms: what we confirmed on the vendor's own pages, and what remains the vendor's word.

Verified: every price, billing period, allowance and per-request cap in the table above, read from the rendered pricing page on all three toggles and cross-checked against the support center's own pricing guide; the $1.00 introductory offer and its footnote; the seat add-on prices in the page's plan configuration; the API's per-word rates, endpoints, models and published Pangram V4 figure; the MCP server address and the integrations marked live; the detector's 200-word no-account scan; the terms on currency, plans and limits, cancellation, "unlimited" use, the money-back policy and the two missing policy sections; the counters on three pages; the university logos; and the status page figures on the day we read them.

Not verified: whether the humanizer evades any detector, at any setting, on any date; the 99% and 89% figures; detector accuracy; the user, company and document counters; whether a re-run deducts words a second time; what the Samurai add-on costs; whether a free humanizer tier exists behind a signup, as one vendor page implies and the pricing page contradicts; what window the status page covers; output quality; latency; and support responsiveness. Each of those is either unstated by the vendor or stated without evidence, and is labeled as such wherever it appears on this page.

That means the score below is a score for pricing transparency, feature breadth, documentation and the trustworthiness of the vendor's own claims. It gives no credit, in either direction, for evasion performance.

The site's claims, side by side

Because the central promise cannot be tested from outside, we read the vendor's pages against each other instead. Every pair below was live on 24 September 2026.

None of this is evidence the product is bad. It is evidence that the marketing runs a long way ahead of what the vendor's own pages support, and that is the surface people buy on.

Where StealthGPT is the strongest pick

StealthGPT is at its strongest as a pipeline component rather than a browser tab. The API is the best-documented part of the product and the cheapest way to buy its words: three REST endpoints under an api-token, $0.20 per 1,000 words on Standard, no subscription, words the page says never expire, and a response returning both a 0 to 100 detection score and the words it spent. A public MCP server at www.stealthgpt.ai/api/mcp/mcp puts the same humanizer inside Claude, Claude Code, Cursor and Manus with no code at all.

The second argument is the allowance at volume: Max on yearly billing is 200,000 words a month for a $41.67 monthly equivalent, near $0.21 per 1,000 words, and Enterprise is a million words for $179.99 a month. The catch is the per-request cap, 1,000 words on Pro and 2,000 on Max whatever your allowance. The third is the Chrome extension, which humanizes highlighted text in place from a right-click inside Google Docs, Gmail, Outlook Web and any other text field, removing the copy-paste loop that makes most tools in this category tedious. It is Chrome only, by the vendor's own answer: "Currently, the extension is available for Chrome only."

Best for

Content and SEO teams with permission to use AI, a real volume of draft copy to clean up, and an automation stack to put the API or the MCP server into. The per-word rates at Max and Enterprise, and the $0.20 per 1,000 words pay-as-you-go tier with non-expiring words, are the reason to choose this over a cheaper entry-level competitor. Not for students or anyone submitting graded work: the vendor publishes no acceptable-use policy, its own FAQ tells students professors cannot tell, and every institution we are aware of treats submitting humanized AI text as misconduct. Not for anyone who needs a detection score they can defend to a client or a panel, a point this vendor's own detector page concedes. And not for a buyer who wants a low-commitment trial: the closest thing on offer is a $1.00 first month that renews at $29.99.

Key features

• Three humanizer models: Super, marketed as built to pass Pangram V4; Standard, described as proven against GPTZero, Winston and Originality; and Lite, for light touch-ups "when detection isn't a concern"
• Four tone levels, Standard, HighSchool, College and PhD, set independently of the model
• A quality mode with built-in checks against a fast single-pass mode, a detail control running High, Medium and Low, and a 0 to 100 detection score returned with every output
• Weekly, monthly and yearly word allowances, with per-request caps of 1,000, 2,000 and 10,000+ words by tier
• Free AI detector with a 200-word Standard Scan that needs no account, and a longer Enhanced Scan that does
• Chrome extension that humanizes highlighted text in place from a right-click in Google Docs, Gmail, Outlook Web and any text field, with custom shortcuts; Chrome only. An iOS app is included in every plan's bullets
• Three REST endpoints under an api-token: /api/stealthify for the writer and humanizer, /api/stealthify/agent for the Stealth Agent workflow, /api/stealthify/balance for usage and remaining words
• Public MCP server, with Claude, Claude Code, Cursor, Manus, n8n, WordPress and Zapier marked live and Codex coming soon; the Zapier connector advertises 8,000+ apps
• Accessory suite: Stealth Writer, Stealth Agent, SEO Writer, Chat with PDF, Photo to Answers, a grammar checker, summarizer, simplifier, sentence rewriter, SEO optimizer, formalizer and a watermark remover with a Claude-specific variant
• Document upload, in-text citations, generation history and a feature the pricing table calls Stealth Insights, listed for every tier
• Public status page at stealthgpt.betteruptime.com, broken out by service; three seats on Max and five on Enterprise, with extras sold as an add-on

Best-fit use cases

• Stripping the recognizable tells out of AI-drafted marketing copy, product descriptions and blog posts where AI use is permitted but robotic phrasing is not
• Bulk humanization inside a content pipeline through the API at $0.20 per 1,000 words that do not expire, or wired into an agent or editor over MCP so a draft is cleaned up in the session it is written in
• Rewriting human-written copy that a client's or an editor's detector has false-flagged, the one use of this category that carries no integrity problem
• Inline cleanup inside Google Docs or a CMS through the Chrome extension, when the copy-paste loop is the bottleneck
• Agency or team use on Max and Enterprise, where seats and a dedicated support channel are included

Pros

• Prices are published in full on three billing periods without a login, and restated in a support-center article that states they are US dollars
• Good per-word economics at volume: roughly $0.25 per 1,000 words on Max monthly, about $0.21 on Max yearly and $0.18 on Enterprise, by our arithmetic from the vendor's own allowances
• A pay-as-you-go API at $0.20 per 1,000 words with no subscription and, by the vendor's statement, words that never expire, removing the monthly reset that defines the rest of this category
• Unusually broad integration surface: three REST endpoints, a public MCP server with Claude, Cursor, n8n and WordPress marked live, Zapier, a Chrome extension and an iOS app
• Every response returns a 0 to 100 detection score and the words spent, and model and tone are separate controls rather than one slider
• A public status page broken out by service, which almost nothing else in this cluster publishes, and a real support center with a knowledge base, a ticket queue, Zendesk and a Discord
• The detector page states the conflict of interest plainly rather than hiding it: the detector is built by the company that bypasses detectors for a living

Cons

• No acceptable-use or ethics policy: the terms list a Usage Policies section and an API Policies section in the contents and the document contains neither
• The marketing is aimed at students explicitly: the humanizer FAQ answers "Can professors tell if I used ChatGPT to write my essay?" with "Not when you use StealthGPT", and the pricing page runs Yale, Johns Hopkins, MIT, Florida and Michigan logos under "Trusted by students from:"
• Section 8 of the terms says all purchases are non-refundable while section 31 publishes a money-back guarantee; that guarantee then excludes anything flagged by Originality.ai, which the same site advertises beating, and does not name Pangram among its acceptable detectors even though the flagship model is sold as built to pass Pangram V4
• Two bypass rates on the same site the same day: 99% on the detector and humanizer pages, 89% against Pangram V4 on the API page, neither with a methodology
• Three language counts (50+ on the pricing counter, 100+ in the plan table on the same page, "over 1,000" in two FAQs) and counters that disagree between pages read the same day (952 companies and 9.4M+ documents on the homepage, 810 and 50M+ on pricing)
• No free plan for the humanizer: the cheapest entry is $6.99 for a week or a $1.00 first month that renews at $29.99, and the detector page's claim that free accounts unlock the humanizer is not what the pricing page sells
• Per-request caps of 1,000 words on Pro and 2,000 on Max mean a long document is several stitched-together requests, and the Pro card advertises seats and priority support the comparison table on the same page marks as not included
• The pricing page's structured data advertises the product at price 0 and publishes a 4.0 rating from 203 ratings whose source is nowhere on the page, next to six five-star testimonials attributed to job titles
• Three corporate addresses in one terms of service, New Jersey, Sheridan in Wyoming and St. Petersburg in Florida, with Wyoming law governing, and a prohibited-activities list that includes "Disparage, tarnish, or otherwise harm, in our opinion, us and/or the Site"
• Whatever evasion rate the tool achieves today is perishable by design and cannot be relied on next quarter

What to watch out for

Four things, in order of how much they should affect your decision.

There is no published usage policy, and the marketing fills the gap. This is the biggest difference between StealthGPT and its closest rival. Undetectable AI publishes an ethics page naming cheating in school as unacceptable use. StealthGPT's terms promise a Usage Policies section and an API Policies section in the contents, then skip from section 9 to section 12. In that vacuum the humanizer FAQ answers "Can professors tell if I used ChatGPT to write my essay?" with "Not when you use StealthGPT", the tone presets run HighSchool, College and PhD, the detector page names students and academics among its audiences, and the pricing page carries five university logos under "Trusted by students from:". A product configured this specifically for coursework, with no policy saying not to, is not neutral about who is buying it. Submitting humanized AI text is academic misconduct at essentially every institution we are aware of, and the risk is entirely the student's.

The guarantee excludes the detectors most likely to catch you. Section 31's "Detected? Get your money back." policy pays out on a screenshot from a valid AI detection service, and refunds weekly and monthly subscribers for the current billing cycle only, annual subscribers the affected month plus a prorated remainder. Then come the carve-outs. It does not apply if you used a passage StealthGPT's own stealth score had already flagged, so the company's self-assessment gates its own guarantee. It does not apply to anything flagged by Originality.ai, dismissed in the terms "due to the invalidity of their AI Detection system" while the same site shows Originality.ai's logo in its detector wall and markets the Standard model as proven against it. The named acceptable detectors are Turnitin, GPTZero, ZeroGPT and Copyleaks, with anything else admitted only if the company deems it precise enough; Pangram is not among them, though Super is sold on an 89% figure against Pangram V4. And section 8 says all purchases are non-refundable. Read both sections before you buy, and do not treat the guarantee as a safety net.

The headline numbers do not survive being read side by side. The bypass rate is 99% on two pages and 89% on a third. The language count is 50+, 100+ or "over 1,000" depending which page you are on. Satisfied companies is 952 on the homepage and 810 on the pricing page; documents processed is 9.4M+ on one and 50M+ on the other, and the API page attaches 50M+ to words instead. The structured data advertises the product at a price of zero beside a 4.0 rating from 203 ratings that appear nowhere on the page. None of this makes the humanizer bad, but when a number on this site is the reason you are buying, check whether another page says something different.

Work out what you are committing to before you click. The $1.00 is a first month, not a trial, and it renews at $29.99 or $49.99. The weekly plan looks cheap at $6.99 but runs $363.48 over a year on Pro, more than the $299.99 yearly price, and the Max yearly card's monthly figure multiplies out to $3.89 above the annual total the vendor's own support guide gives. The per-request cap, not the allowance, is what will slow you down: 1,000 words at a time on Pro.

Is StealthGPT worth it?

Yes, for a narrow buyer: a content or SEO operation with permission to use AI, real volume, and somewhere to put an API key. The per-word economics are good, the $0.20 per 1,000 words pay-as-you-go tier with non-expiring words is the most sensibly priced thing the company sells, and the MCP server and Zapier connector let the humanizer live inside a pipeline rather than a browser tab.

It scores 3.4 rather than higher for reasons that have almost nothing to do with output quality. The product surface is broad and the documentation beats the category average. What holds the score down is the company's own material: no acceptable-use policy while running university logos above a student-facing pitch, a refund guarantee that contradicts a non-refundable clause in the same document and excludes a detector the site advertises beating, a bypass rate that is 99% on two pages and 89% on a third, a language count with three values, and structured data that prices the product at zero. A vendor this precise about word allowances and this loose about everything else is telling you which claims it expects to be checked.

Worth saying plainly: whether the humanizer beats Turnitin, GPTZero or Pangram this month is the question most readers arrived with, and this page does not answer it. Nobody can, durably. Our detector guide cites University of Chicago research that found the strongest detector kept catching StealthGPT-rewritten text while weaker ones missed more than half of it, and StealthGPT now markets a model built specifically to pass that detector's fourth version. That is the arms race, and it will have moved again by the time you read this.

How we arrived at 3.4

The score is weighted across eight categories, each marked out of 5. The weights are published so the number can be reproduced, and value and market trust carry the most weight because they are the two things a buyer of this product can evaluate from outside.

The 5 point score is a direct weighted mean of the table, and it reflects a well-integrated, fairly priced humanizer sold by a vendor whose own pages are the strongest argument for reading the terms before you buy.

FAQ

Pricing, billing periods, plan limits, add-on prices, API rates, integration list, refund terms and feature claims verified 24 September 2026 from stealthgpt.ai/pricing (rendered on the weekly, monthly and yearly toggles, plus the page's underlying plan configuration and structured data), the support center's "Current plans and pricing" guide, stealthgpt.ai/stealthapi, stealthgpt.ai/free-ai-detector, stealthgpt.ai/feature/ai-humanizer, stealthgpt.ai/feature/chrome-extension, stealthgpt.ai/tos, the homepage and stealthgpt.betteruptime.com. All prices are US dollars, which the terms of service state directly. StealthGPT is scored against our published 8-point framework; every bypass rate on this page is the vendor's own claim and we make no independent evasion claim. ToolChase received no payment for this review. Scores, rankings and list positions are never sold. See our methodology for how we score tools.