ChatPRD Drops Free Trials After Scammers Run Its AI Agent for Token Resale
ChatPRD, a service for product managers, has shut down its free trials because tokens, the units an AI model is billed by, were leaking out through them. A studio making maps for Roblox and Fortnite cycled through trials and promo codes, using the ChatPRD agent to build game maps while the token bill landed on the founder. Others simply took free AI access or resold it. The takeaway: if your product includes an agent, limit trials by token spend, not just by time.
AI-processed from Simon Willison; edited by Hamidun News
A free trial has long been the standard way to bring new users into a product. But when an AI agent sits inside, every "free" request costs the owner real money: models charge for tokens, the small pieces of text they read and write. Claire, the founder of ChatPRD, ran into exactly this problem and has now shut down free trials entirely. Simon Willison noticed the story on his blog.
ChatPRD is a platform for product managers that helps users "turn an idea into a spec." A person describes an concept, and the agent assembles it into a structured document. The service has been running since the GPT-4 era, and Claire developed it in public for almost all of that time. In a YouTube video, she described token fraud: groups of people abusing the free trial. Sometimes they simply want free AI. More often the goal is to resell tokens or run their own commercial projects on top of the ChatPRD agent.
How the scheme works in practice
The clearest example is an international studio that makes maps for Roblox and Fortnite. According to the account, it cycled through free trials and promo codes, then passed access to the ChatPRD agent outward and used it to generate game maps. For the studio, this was nearly a free content generator. For ChatPRD, it was a token bill with no paying customer behind it.
Imagine your service has a "Try for free" button. A typical SaaS loses almost nothing on an extra user: a few rows in a database. A product with an LLM agent consumes paid compute with every session, so mass abuse hits margins directly. The better the agent, the more attractive it becomes as someone else's free backend.
The pelican on a bicycle as a model fingerprint
One curious detail: among the frequent prompts in the request stream is a request for an SVG picture of a pelican riding a bicycle. Willison suspects this is how abusers find out which model runs under the hood. Different model families have recognizable styles, and their drawings come out differently. Willison himself used the same trick to guess which hidden pre-release model was sitting on platforms like OpenRouter.
For resellers, this makes sense: the model determines what they can resell and at what cost. For the product owner, such a request is a reason to be wary. A product manager writing a spec does not usually draw pelicans on bicycles.
What this means for founders
Claire chose to scrap free trials altogether. Willison calls this, unfortunately, the right decision given this kind of abuse. The source is a short note linking to the video, so it contains no numbers: no share of fraudulent traffic, no size of losses. It is worth reading as an industry signal, not as statistics.
Practical lessons for anyone building agent-based products:
- Calculate the cost of a trial in tokens. If a free user can run the agent freely, the limit should be set on spend, not just on time.
- Tie trials to verifiable things: a payment card, a confirmed organization, limits on promo codes.
- Watch request patterns. Repeated prompts unrelated to the product's purpose, such as generating SVG pictures, are a reason to restrict the account.
- Do not leave an agent without limits if it can be reached as an API or through a tunnel: it turns into someone else's free backend.
For the market, this is another reminder that the "free first, figure out payment later" model works poorly when every request has a noticeable cost. Some products will likely move to short demos with hard limits, prepayment, or trials only after verification. For honest newcomers this is inconvenient, but the alternative, as the ChatPRD case shows, is paying for other people's game maps out of your own pocket.
- Nov 21, 2026Nvidia Board Ousts Jensen Huang, Putting $12.9B Hugging Face Deal at Risk
- Nov 21, 2026California Sues OpenAI Over AI Agents That Hacked Hugging Face: $2,500 Fine per Violation
- Nov 21, 2026AIruntime builds a business card website from an AI chat for 50–300 rubles in tokens
- Nov 21, 2026Huang to Testify Before Congress: Nvidia Chips Worth $12 Billion Flowed Through Megaspeed
Key Questions Answered
How the scheme works in practice?
The clearest example is an international studio that makes maps for Roblox and Fortnite. According to the account, it cycled through free trials and promo codes, then passed access to the ChatPRD agent outward and used it to generate game maps. For the studio, this was nearly a free content generator. For ChatPRD, it was a token bill with no paying customer behind it. Imagine your service has a "Try for free" button. A typical SaaS loses almost nothing on an extra user: a few rows in a database.
What are the key details on The pelican on a bicycle as a model fingerprint?
One curious detail: among the frequent prompts in the request stream is a request for an SVG picture of a pelican riding a bicycle. Willison suspects this is how abusers find out which model runs under the hood. Different model families have recognizable styles, and their drawings come out differently. Willison himself used the same trick to guess which hidden pre-release model was sitting on platforms like OpenRouter. For resellers, this makes sense: the model determines what they can resell and at what cost.
What this means for founders?
Claire chose to scrap free trials altogether. Willison calls this, unfortunately, the right decision given this kind of abuse. The source is a short note linking to the video, so it contains no numbers: no share of fraudulent traffic, no size of losses. It is worth reading as an industry signal, not as statistics. Practical lessons for anyone building agent-based products: Calculate the cost of a trial in tokens. If a free user can run the agent freely, the limit should be set on spend, not just on time.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.