The third attempt followed two small failures
Dinh's own newsletter says TypingMind was his third attempt at building an AI product. The previous two, EmojiAI and AskCommand, did not work out and earned roughly $100 together. That record is narrower and more credible than calling the story six failed AI projects.
When Dinh left his job, his public account listed about $300 in monthly recurring revenue from one product and about $200 per month from another. He also had savings, product skills, and a growing audience. Those conditions created room to experiment, but they do not turn the later result into a repeatable shortcut.
The first version shipped five days after the API release
Dinh started with problems he experienced in the existing chat interface, including weak chat history and a frustrating user experience. He registered the domain, worked on the prototype, and released the first version five days after the API announcement. Attention came first; he then added payment and received an initial sale.
The editorial interview records his founder-reported progression from $1,000 in revenue to $2,000, $4,000, and $10,000. A Product Hunt launch, supported by his existing Twitter audience and newsletter, reportedly moved cumulative revenue from $10,000 to $22,000 in one day.
One-time licensing matched the original cost structure
TypingMind was initially a static web app without a backend, server, or database. Customers brought their own API keys and stored them locally. With little recurring infrastructure cost for the founder, a one-time license could cover access without requiring a monthly subscription for every individual buyer.
This structure did not remove business risk. One-time sales fluctuate, continuous marketing is needed to reach new customers, and new features must remain useful. Platform changes and competition also remain serious risks. A low recurring cost base explains the pricing choice; it does not guarantee durable revenue.
The audience and launch window cannot be copied
Dinh already had experience shipping multiple products, an established Twitter and newsletter audience, and the benefit of an early model API release window. Product Hunt distribution amplified that position. Removing those factors would make the sales timeline misleading.
A new builder should translate the case into a smaller question: is one person willing to pay once for a specific interface improvement? That preserves the useful mechanism while discarding the assumption that another founder can reproduce the same attention, timing, or cumulative skills.
Sell one interface improvement before expanding
Choose one interface problem you encounter frequently. Define one user, one input, one output, and one price. Show the result to a potential buyer before adding a broad feature set. If no stranger will pay, change the problem or the offer instead of treating free registrations as proof of demand.
Within the models currently and lawfully offered at https://APIToken.Company, check the public status page, create a separate project API key, and run one small real call. Record usage, failures, retries, and human correction time under a stopping budget. A visible model is not automatically suitable for a task, and one successful request is not production readiness.
Keep the evidence and revenue labels visible
The sources are an Indie Hackers editorial interview and Dinh's founder newsletter. The $22,000 figure is founder-reported cumulative one-time sales by the cited launch window. It is not monthly recurring revenue, net profit, audited cash receipts, or a result promised to another founder.
The public accounts do not provide a complete refund, tax, support, or labor-cost ledger. They also provide no evidence that Dinh or TypingMind used APIToken. APIToken is presented only as a controlled way to isolate a key, observe status and usage, and bound a small model test.