Allegations of Deceptive Routing and Fraud Surface Against CrofAI
CrofAI, an artificial intelligence inference provider that marketed itself as the world’s most affordable provider of large language models, is facing serious allegations of deceptive business practices. According to reports circulating on Reddit and investigative findings published by developer Kendell, the company allegedly functioned as an “OpenRouter wrapper,” routing user requests to significantly cheaper and weaker models than those advertised.
Discrepancies in Model Routing and Pricing
The core of the controversy involves a massive markup on token pricing. Investigations suggest that CrofAI sold high-end models at premium rates while secretly serving inferior alternatives. For instance, while the company charged between $2 and $10 per million tokens for the kimi-k3 model, it allegedly routed those requests to GLM 5.3 Flash via OpenRouter. This represents a 13.3x markup on input tokens and a 20x markup on output tokens.
Furthermore, CrofAI’s proprietary model families—including greg-2-ultra, greg-1-mini, and greg-2-super—were allegedly found to be aliases for existing models such as GLM 5.2, Qwen 3.5 9B, and Kimi K2.7 Code. According to unverified reports on Reddit, the company’s owner admitted in private messages that claims regarding the originality of the “greg” model family were untrue.
Technical inconsistencies have also been raised regarding the company’s hardware claims. CrofAI asserted it was running Kimi K3 on rented RTX Pro 6000 GPUs via Vast; however, analysts noted that the model’s memory requirements exceed the capacity of the largest available hardware configurations on that platform. Additionally, the company claimed to run deepseek-v4-flash-0731 on a local DGX Spark, a machine with only 128GB of memory, which is insufficient for the model in question.
Failed Remediation and Disappearance
Following the initial exposé, CrofAI attempted several technical adjustments to hide the “fingerprints” of OpenRouter, though investigators claim these attempts failed to resolve the underlying routing issues.
The company’s response to the mounting allegations has been characterized by rapid shifts in communication. After initially denying the allegations, the company issued a statement via a blog post—purportedly written by a “team” rather than the founder—claiming the founder had acted under extreme stress and announcing a service resumption in two weeks. This was accompanied by a social media update where the account owner, identified as “Nathan,” claimed to be stepping back to allow a new team to take over.
However, this transition appears to have been short-lived. Within hours of these announcements, CrofAI wiped its entire digital presence. The company’s primary domains, crof.ai and nahcrof.com, now return 404 errors, the official Twitter account has been deleted, and the company’s subreddit has been set to private.
Impact on Users
The sudden shutdown has left users seeking recourse. While some users reported receiving refunds for unused credits, others have noted that credits used for the supposedly “expensive” models were not reimbursed. The incident has prompted discussions within the developer community regarding the lack of regulation in the rapidly evolving AI inference market.