How it runs

Same prompt for every model. One run. Up to 60 minutes and $0.60. The file is on the page.

One run

Each model gets one run per prompt. The first run that produces the file is the one on the site, and nobody picks the best of several tries. A run that ends without the file, or runs out of time, is run once more. If there is still no file, that model is left off that prompt, and the wall says so.

Runner, network and provider errors don't count as a run. Those runs are discarded and run again.

The daily job

A job runs every day at 05:17 UTC. It reads the OpenRouter model list and adds text and image models released in the last 21 days that cost $25 or less per million output tokens (free routes always qualify). It then runs the prompts, free routes first and paid models newest first, four paid models at most per pass. After that it takes stills, looks up Hacker News and Hugging Face, rebuilds this site and uploads it.

Money

Each API run is held to a $0.60 ceiling: max_tokens is the smallest of the route's limit, the model's own maximum and $0.60 divided by the output price. The job starts a paid run only when the last 7 days of API runs stay within $3.00 and the account's credit stays above $1, counting this run and every paid run already going at its worst case. When a run ends, its real cost replaces its worst case. This week so far: $0.00 of $3.00. Free routes run first and cost nothing. The 21 Oct launch backfill spent $4.13 and does not count toward the weekly budget.

The cost on a row or a poster is the list price: the run's tokens times the model's price per million on OpenRouter. A run with no known list price shows no cost. The run page also shows what was billed.

Same prompt for every model

The prompt is the task plus one fixed line saying which file to deliver. Every run page shows the full prompt.

Draw a mechanical wristwatch that shows the current local time, with a sweeping seconds hand and a date window. Make it look expensive.

It must be one complete, self-contained HTML file: inline all CSS and JavaScript, and make no network requests of any kind (no CDNs, web fonts or remote images). Reply with the file in a single ```html code block.

Six prompts go to every text model, and one image prompt goes to the image models, without the delivery line.

Default settings

Models run on their default settings with nothing added: no custom system prompt, skills, plugins or memory of earlier runs. API runs are a single user message to OpenRouter. Each run gets the model's own maximum output, or less when the $0.60 ceiling is lower.

63 older runs went through the vendor's own coding agent instead (Claude Code, Command Code, opencode). That agent brings its own system prompt and tools. Each run page names the route.

Pages run with the network off

The prompt asks for one self-contained file. Pages are shown in a sandbox that blocks every outside request. A page that depends on a CDN or a remote image is shown without it, and the run page lists what it tried to load. Browser storage and cookies work in memory for the length of your visit, so a game can still keep a high score. The run page shows the file exactly as the model wrote it.

Time and order

Time is wall-clock time from sending the prompt to receiving the file. Walls and rows are ordered by the model's release date, newest first.

Questions

Why one run?

Every run costs money and time. A model gets the file it produced the first time, and a run without a file is retried once.

Why is my model missing?

It may be queued for the next daily run, it may be older than 21 days, it may cost more than $25 per million output tokens, or its output limit may be too short for a page. It may have no OpenRouter route, or this week's $3.00 may be spent. Models without runs are on the Models page under "Not run yet", with the reason.

What is the difference between API and CLI runs?

API runs are one user message to OpenRouter with no system prompt. CLI runs went through the vendor's coding agent with its own system prompt and tools. The numbers line under every output ends in API or CLI.

Who pays for the runs?

The site's owner, from one OpenRouter account capped at $3.00 of runs a week. Free routes cost nothing, and the CLI runs used existing subscriptions.

What does the HN number count?

The points of Hacker News stories whose title matches the model's name, from 3 days before its release until the last build, added up. The link opens that exact search on Algolia. "no HN match" means the search found no story.

Can I suggest a model?

Every new model on OpenRouter is picked up by the daily job, and the RSS feed lists new ones. Models that are only reachable outside OpenRouter are not covered yet.