Tony Dinh drops $124/month i18n SaaS for a $0.75 run

Tony Dinh cancelled the translation SaaS he was paying $124 a month for and moved internationalization into what he calls a built-in translation harness. He ran 2,705 translation keys across 16 locales for about $0.75 with GPT-Luna 6.

Tony Dinh

@tdinh_me

X

2,705 translation keys x 16 locales. Costs ~$0.75 with GPT-Luna 6

The interesting part is not the price, it is what he feeds the model. The harness passes app context along with each string, specifically a list of internal words the app uses, where the string appears in the product, which feature and which screen, and how much screen space the text has including whether it can wrap. That last one is the detail a generic translation pipeline cannot know and the reason translated UI strings usually overflow a button.

What the number is and is not

The $0.75 is a single batch run over 2,705 keys times 16 locales, not a monthly bill. The $124 was recurring. Those two numbers are not directly comparable unless you know how often the key set changes, and Dinh did not post a re-translation cadence. A product that adds strings every sprint pays the batch cost again each time, though at this unit price the comparison still looks one sided.

The gap in the post is accuracy. Dinh says the output is "much more accurate AND in-context" than what the SaaS produced, but he published no eval, no sample of disputed strings, and no native speaker review across the 16 locales. That is one builder's judgment of his own app's output, not a measurement. Translation quality also degrades unevenly by language, so a result that reads fine in the locales you can personally check tells you little about the ones you cannot.

For people building on it

The reusable idea here is the context payload, not the model choice. Most i18n tooling treats a string as a string, which is why translators produce text that is correct and still breaks the layout. Attaching feature, screen and space constraints to each key is work you do once in your own codebase, and it is the part a vendor cannot do for you because the vendor does not have your screens.

If you try this, budget for the review pass Dinh did not describe. The cost of the generation is clearly small. The cost of shipping a wrong string into a paid product in a language nobody on the team reads is not.