网页 综合 更新时间 2026-08-02 19:18:23

Netlify's Hot AR Summer Challenge: I Wasn't Going to Enter. 33 Apps Later, I'm So Glad I Did - DEV Community

自动蜘蛛 dev.to 正文约 5999 字
Netlify's Hot AR Summer Challenge: I Wasn't Going to Enter. 33 Apps Later, I'm So Glad I Did - DEV Community
Posted on Aug 2 • Edited on Aug 2
On Friday, July 31, I was at date night dinner with my wife in LA. We sat down at 8. The plan was simple: enjoy dinner, drive the hour home by 11, post my last app, close out the month.
It was a five course tasting menu. At 9:15 we had finished exactly one course.
I did the math at the table. Four courses to go, an hour of driving after that, and a deadline that did not care about either. So I picked up my phone in panic. Not my laptop, which was at home. My phone.
I built the last one right there at the table, sitting across from my wife, between courses. Claude for the build. Then I deployed it from the same phone through Netlify Agent Runners running Gemini. A confetti canvas, a terminal card, a certified stamp, and a deterministic serial number, because the closing piece had to prove its own work too. Last app of the month!
It went live. On the first try. From a restaurant table.
Then I went to submit it and found there was nowhere to put it. The window had closed. My stomach dropped somewhere around the second course.
We left at 11:30 and got home at 12:30 in the morning. Thirty-three apps made the wall. The thirty-fourth exists, works, and was never counted.
That is the ending, so I am going to start there and work backward.
Netlify ran a building challenge called Hot AR Summer through July. Build with Agent Runners, ship a small app, put it on the wall. Two apps a day allowed. I saw it and thought, that is a nice idea for people with more time than me.
Then I built one. Then the next day I built another one. Then it stopped being a decision.
By the end I had 33 apps on my builder page, out of a showcase somewhere north of 200 total from everyone who played. A worry filing system with real Dewey Decimal call numbers. A diner wheel that decides dinner by how long you hold the button. A quiz that tells you which dead language you would think in. A goth subtype sorter I built for my wife. A basketball dashboard about Luka. A coin flip that shows its arithmetic.
None of them are big so I could move fast and break things. Ha! One screen, one idea, one self-contained index.html, one deploy.
What I did not expect was that the constraint I picked for aesthetic reasons would turn out to be a budget decision.
Here is the rule I ran all month: no Math.random in any selection logic.
Every app decides its output from the input. Type the same worry into Dewey, get the same call number. Hold the dinner wheel for the same 1.8 seconds, get the same supper. Press the Honest Coin at the same millisecond, get the same face and the same hash to prove it. The apps hash their inputs, run a fixed rule table, and show you the derivation.
I like this for philosophical reasons. I came to code from the courtroom, and in that world a confident wrong answer is worse than no answer. A receipt you can check beats a result you have to trust.
But the reason it survived contact with a credit meter is plainer than that.
A deterministic app calls a model zero times after it is built.
All of the AI cost is in construction. Once the thing is on Netlify it is a static file doing arithmetic in the visitor's browser. There is no inference at runtime. There is no API key. There is nothing to rate limit, nothing to cache, nothing that gets more expensive when more people show up.
I did not fully appreciate what that meant until I opened the billing page.
Across July I paid $44 in top-ups. My current billing period shows 3,990 credits consumed. Here is where every one of them went.
That works out to about 121 credits per app, or roughly $1.33 an app if you divide the top-ups by 33.
Three quarters of my spend was the model reading and writing code. Another quarter was deploying and the machine time to do it. And then there is that fourth row, which I want to sit on for a second.
3,160 web requests. Six tenths of one credit.
Sixty percent of a single credit, to serve every visit to every app I shipped, for a month. That is not a rounding error in my bill, it is the rounding error. Serving the apps cost less than one twentieth of what a single deploy cost.
Here is the counterfactual that made me put my coffee down. Suppose I had built these the usual way (i.e. me) a lot of "AI app" demos get built, with a model call on every visit. Some little inference to generate the flavor text instead of computing it.
Using my own measured Gemini rate from this month (2,254.97 credits for 16M tokens) and a conservative small call of about 1,200 tokens per visit, those same 3,160 visits would have run roughly 534 credits. Against the 0.6 I actually paid.
That is about 890 times more expensive, and it would have cost nearly as much as all 41 of my deploys combined. At my actual traffic. Which is small. That number does not stay still if anything ever catches on.
The 0.6 credits is measured. The 534 is my estimate, using my own token pricing and an assumed call size, so argue with it if you want.
Either way the difference is not really about my traffic, it is about what happens next. My apps cost the same to run whether ten people visit or ten thousand. An app that calls a model on every visit gets more expensive with every single visitor. One is a flat line. The other climbs forever. I got the flat line for free by being stubborn about Math.random.
I ran two providers through Agent Runners this month, and the split is instructive.
OpenAI cost me about 1.8x per token for the same category of work: generate a single-file HTML toy. It is not that it did a bad job. Scrivener, my DOS-terminal typing tutor, came out of Codex and it is one of the better-built things on the wall. It just is not 1.8x better at making a small deterministic app, and when you are shipping one a day that multiple compounds.
So the rule I landed on mid-month, and wrote into my tracker so I would stop relitigating it: default to Gemini, sketch it out by hand first, one deploy, do not re-prompt a working app.
That last clause is the one that
访问原文 返回搜索 https://dev.to/earlgreyhot1701d/netlifys-hot-ar-summer-i-wasnt-going-to-enter-33-apps-later-heres-the-bill-1d25
本页为82搜索生成的网页快照,内容来自站内索引;演示数据的原文链接仅作示例。