Habr AI compared AI rewriters: many services proved to be ChatGPT wrappers

The market for AI rewriting services targeting Russian-language newsrooms turned out to be far less technologically sophisticated than its packaging suggests. A comprehensive analysis on Habr AI shows that a significant portion of services either continues to solve the old SEO task—raising uniqueness percentage for the sake of a report—or simply sells access to the same LLM through a more convenient interface with a noticeable markup. True editorial quality today comes not from "magic buttons," but from properly configured models, prompt libraries, and multi-agent pipelines with several stages of text review.

The analysis covers 24 tools, which the author divides into four major categories. The first category consists of old-school SEO rewriters like Text.ru, Advego, Raskruty, and ETXT.

These services mostly rearrange words and select synonyms to achieve the desired uniqueness percentage, but barely work with meaning. The second category includes general LLMs: Claude 4.6 and 4.

7, GPT-5, Gemini 3, DeepSeek V3, YandexGPT 5, and GigaChat. They are noticeably stronger in text quality, but without a good prompt, they often still produce an averaged "like from a chatbot" result. The third group consists of ready-made wrappers based on the same models, such as PR-CY, Turbotext, Chad, Gerwin AI, and NeuralWriter.

Their value most often lies in the interface, templates, and pricing, rather than proprietary technology. Finally, the fourth category includes multi-agent and vertical pipelines, where several models or agents sequentially gather facts, write a draft, critique it, and refine it to a final version. To make the comparison not merely theoretical, the author ran one short news article of approximately 1500 characters through several services with a simple request to "rewrite this text in your own words."

In this test, Claude Sonnet 4.6 received the best subjective score—7.5 out of 10, showing minimal clichés and zero hallucinations.

GPT-5 came in slightly lower with a score of 7 out of 10. DeepSeek V3 looked strong in terms of price-to-quality ratio but made a factual error and inserted an extra digit. YandexGPT 5 and PR-CY showed average results, while the old synonym replacer Raskruty essentially just rearranged words and scored 2 out of 10.

It is separately emphasized that none of the products in the selection provides out-of-the-box fact-checking at the level of a live editor, so human verification remains mandatory. The review also draws a practical fork in usage scenarios. If newsrooms need to rewrite one or two articles per month, it is usually more cost-effective to go directly to Claude or ChatGPT and spend time once on a proper prompt.

If the flow reaches 10–30 materials per month, maintaining personal prompt libraries in Notion or Word, as well as convenient LLM wrappers, start to pay for themselves. If the task is to consistently publish 30 or more texts in the style of a specific publication, with background context, fact-checking, and draft refinement, then multi-agent solutions come to the forefront. They are more expensive and complex, but editorial quality today is built around the process, not around a single model.

For companies for whom it is critical to store data within the Russian perimeter and comply with the requirements of Law 152-FZ, YandexGPT and GigaChat can be a separate argument, although in terms of quality and freedom to work with the agenda, they fall short of Western models. The main thesis of the material is that in 2026, the word "rewrite" conceals several different tasks. This could be meaningful reworking of news to match a specific media outlet's style, adapting the same story for Telegram, Zen, or a long-form article, or mechanical text uniqueness for SEO checks.

When a customer mixes these tasks into one, they almost inevitably overpay or receive poor text. The author separately notes that the comparison scale is subjective and based on one test text; moreover, the multi-agent category mentions the author's own product. Therefore, the review is useful not as a final ranking of winners, but as a market map: old SEO tools still live by virtue of uniqueness requirements, direct access to strong LLMs remains the most flexible option, and true differentiation begins where an editorial process appears on top of the model.

For media, this means a simple thing: it makes sense to pay for an "AI rewriter" only when the service adds real value—style control, convenient workflows, data restrictions, or multi-step text processing. If everything inside is the same ChatGPT or another mass-market model without serious augmentation, direct access to an LLM will almost always be cheaper and more flexible. The market bottleneck now lies not in text generation itself, but in how to turn a draft into reliable material that is not embarrassing to publish under the media outlet's name.