{"slug": "stealthgpt-ships-super-after-saying-pangram-broke-its-previous-models", "title": "StealthGPT ships Super after saying Pangram broke its previous models", "summary": "StealthGPT founder Jozef Gherman announced Super on September 8, claiming the new text-rewriting model evades AI detection after Pangram 4 broke its previous models. In StealthGPT's internal test, Pangram V4 classified 89 of 100 Super rewrites as human, but Super failed StealthGPT's factual-consistency check on 18% of outputs, scoring 82% overall. Super is available via StealthGPT's website, REST API, and Model Context Protocol integration.", "body_md": "# StealthGPT ships Super after saying Pangram broke its previous models\n\n**Founder Jozef Gherman's internal test says 89 of 100 rewrites passed Pangram V4, while 18% failed StealthGPT's factual-consistency check.**\n\n        By [RuntimeWire Staff](/author/runtimewire-staff)\n        · Published \n\nPrimary source: [StealthGPT](https://x.com/stealthgpt/status/2097325174861266973)\n\n## Why it matters\n\nAI detectors and humanizers are training against one another in a continuing cycle. Super's claimed bypass rate is notable, but its 82% factual-consistency score shows the tradeoff.\n\n[StealthGPT](https://www.stealthgpt.ai/?ref=runtimewire) founder [Jozef Gherman](https://jozefgherman.substack.com/p/introducing-super?r=gb0zv&ref=runtimewire) announced Super on September 8 after saying Pangram 4 broke StealthGPT's previous text-rewriting models.\n\nThe launch puts Gherman back inside the contest he bet on when he started StealthGPT in February 2023: detection models improve, humanizers adapt, and each release creates the next product requirement. In his launch post, Gherman claimed Super can make AI-generated writing less likely to be identified by leading detection systems.\n\nSuper is available through StealthGPT's website, REST API and [Model Context Protocol integration](https://www.stealthgpt.ai/mcp?ref=runtimewire).\n\nGherman is a self-taught developer who began coding in 2020 because he wanted to build RuneScape bots. He learned through YouTube tutorials and Harvard's CS50 course, then used ChatGPT as a programming tutor, according to a [2024 Indie Hackers interview](https://www.indiehackers.com/post/growing-from-0-to-186k-mrr-in-15-months-fgJbWfKrigaJkqkTdVmg?ref=runtimewire). He built StealthGPT's first version with Next.js, React, Tailwind and a $250 loan to the business while taking consulting work.\n\n### An 89% result from StealthGPT's own test\n\nSuper's headline number comes from an internal benchmark rather than an independent evaluator. In the [published test repository](https://github.com/StealthGPT-Labs/humanizer-benchmark/tree/main/reports/2026-09-07-stealthgpt-super?ref=runtimewire), StealthGPT says Pangram V4 classified 89 of 100 Super rewrites as human. Two received a mixed classification and nine were labeled AI.\n\nThe report also gives claimed bypass rates of 97% for GPTZero, 92% for Winston AI and 94% for Originality.ai. The repository labels the run September 7, while Gherman's announcement says the samples were tested September 6.\n\nThe test establishes that Super performed well on StealthGPT's sample under StealthGPT's procedure. It does not establish an 89% bypass rate across other datasets, writing styles, document lengths or future versions of Pangram. The sample source and selection process also matter because detector performance can change with topic, language, editing history and the model that generated the original text.\n\n[StealthGPT's published quality evaluation](https://github.com/StealthGPT-Labs/humanizer-benchmark/tree/main/reports/2026-09-07-stealthgpt-super?ref=runtimewire) gave Super an 82% pass rate for factual consistency. That means 18 of the 100 tested outputs failed StealthGPT's factual-consistency check, even as most avoided an AI classification.\n\n[Gherman wrote in his announcement](https://jozefgherman.substack.com/p/introducing-super?r=gb0zv&ref=runtimewire) that a rewrite can sound natural and evade detection while changing claims, numbers or meaning. StealthGPT says Super uses a multistage pipeline to inspect structure, factual consistency, completeness, naturalness, stance, sentence integrity and omissions, then checks the output again after making fixes. The 82% result shows that process still has room to improve.\n\n### Pangram forced the rebuild\n\n[Pangram, an AI-detection startup, introduced V4](https://www.pangram.com/blog/introducing-pangram-4?ref=runtimewire) on July 29 with an internal benchmark claiming it identified AI involvement in 98.83% of samples processed by 13 commercial humanizers. Pangram says V4 was built to detect humanizer programs and adversarial prompting.\n\n[Gherman wrote in his Super announcement that Pangram V4 \"broke our previous models.\"](https://jozefgherman.substack.com/p/introducing-super?r=gb0zv&ref=runtimewire) The admission is useful context for the launch: Super is a direct reaction to a detector upgrade, rather than a routine model refresh dressed up as one.\n\nThe cycle also explains why neither side's benchmark settles the matter. Pangram's 98.83% result came from Pangram's dataset. StealthGPT's 89% bypass result came from a separate StealthGPT dataset. Pangram has not independently replicated the Super result, and the published materials do not include a neutral head-to-head evaluation using an agreed test set.\n\nPangram has financial backing to keep the contest moving. Pangram [raised $9 million alongside the V4 launch](https://techcrunch.com/2026/07/29/as-ai-content-floods-the-internet-pangram-raises-9m-to-detect-it/?ref=runtimewire), while Gherman describes StealthGPT as bootstrapped with personal savings and friends-and-family capital. For StealthGPT, each detector update can threaten the core product and create a reason for existing customers to pay for the replacement.\n\n### Gherman's product is moving into content infrastructure\n\nAs of September 9, 2026, StealthGPT's [API page](https://www.stealthgpt.ai/stealthapi?ref=runtimewire) lists more than 700 API users, over 350 businesses and more than 50 million generated words, all company-reported figures.\n\nSuper's API and MCP availability turns detector evasion into a programmable step that can sit inside larger content pipelines using tools such as Claude, Cursor, n8n, WordPress and Zapier. For Gherman, the implication of API and MCP distribution is that StealthGPT can move beyond individual users pasting essays into a web form. Automated integrations could process greater volumes and make the humanizer a recurring step in a production workflow, though StealthGPT has not disclosed customer usage data establishing that pattern.\n\nAs of September 9, 2026, StealthGPT's [API documentation](https://docs.stealthgpt.ai/?ref=runtimewire) lists Standard and Lite at the broader developer rate of $0.20 per 1,000 words. Super is priced separately at $0.05 per 100 words, equivalent to $0.50 per 1,000 words.\n\nGherman's earlier career includes J Squared Invest, a digital-asset investment operation he co-founded and led. In 2021, the [Commodity Futures Trading Commission issued an enforcement order alleging](https://www.cftc.gov/PressRoom/PressReleases/8381-21?ref=runtimewire) that Gherman and J Squared made misleading statements or omitted material facts while soliciting more than $300,000 from over 40 customers. According to the CFTC, the order required $247,110 in restitution, a $150,000 civil monetary penalty and a 10-year ban from activities requiring CFTC registration.\n\nGherman subsequently built StealthGPT around a durable technical and social conflict. Generative writing keeps becoming easier to produce, while schools, publishers and software platforms keep adding systems intended to identify it. Super is his latest attempt to stay one release ahead. Its internal detector scores show progress against Pangram V4; StealthGPT's own factual-consistency result shows the cost of treating evasion as the primary benchmark.", "url": "https://wpnews.pro/news/stealthgpt-ships-super-after-saying-pangram-broke-its-previous-models", "canonical_source": "https://runtimewire.com/article/stealthgpt-super-pangram-ai-detector-bypass", "published_at": "2026-09-09 06:39:48+00:00", "updated_at": "2026-09-09 06:57:43.582966+00:00", "lang": "en", "topics": ["generative-ai", "ai-products", "ai-tools"], "entities": ["StealthGPT", "Jozef Gherman", "Pangram", "Pangram V4", "GPTZero", "Winston AI", "Originality.ai"], "alternates": {"html": "https://wpnews.pro/news/stealthgpt-ships-super-after-saying-pangram-broke-its-previous-models", "markdown": "https://wpnews.pro/news/stealthgpt-ships-super-after-saying-pangram-broke-its-previous-models.md", "text": "https://wpnews.pro/news/stealthgpt-ships-super-after-saying-pangram-broke-its-previous-models.txt", "jsonld": "https://wpnews.pro/news/stealthgpt-ships-super-after-saying-pangram-broke-its-previous-models.jsonld"}}