We may earn money or products from the companies mentioned in this post.
Eight hours. One prompt. About $550 at API rates. That’s how Anthropic engineer Felix Rieseberg described the initial run behind Saltline, a first-person shooter that runs in a browser. He used Claude Fable 5.2 and asked for a small level with ambitious visuals. The reported cost makes this an unusually concrete example of what a long AI coding session can produce.
Open the original one-prompt build or open the later revision. If you’re judging what Claude produced during the eight-hour run, start with the original.
What was playable after eight hours?
Saltline takes place in a rain-soaked container yard. You move through the port, deal with robotic enemies, and activate beacons. It’s a single playable mission with an objective and opposition. That’s a better measure of the experiment than whether a screenshot looks ready for a store page.
Rieseberg’s prompt combined two instructions that are easy to separate. He asked for high visual detail and kept the game itself small. A compact level gives an agent fewer mechanics and locations to connect. It also gives the person reviewing the build an obvious task: get through the level and see whether the core interaction holds up from start to finish.
The demo is also unusual for how it handles visuals. Coverage of the build describes its visuals and models as generated in code, with shaders doing part of the work. That means a realistic-looking asset in the game needn’t have started as a conventional image file or downloadable 3D model. The TweakTown write-up has footage and links to both versions.

What did the $550 pay for?
| Reported figure | What it describes |
|---|---|
| About 8 hours | The run that produced the first playable build. |
| About $550 at API rates | Rieseberg’s estimate for that specific run. |
| About 445 million input tokens | Reported aggregate input, with heavy caching. |
| About 2.7 million output tokens | Reported output over the run. |
Rieseberg gave the figures in his November 9 post. The cache detail matters because repeated input and newly processed input can be priced differently. Multiplying the entire reported input volume by the standard uncached rate would give a misleading estimate of this run. The 445 million figure describes aggregate usage, including repeatedly made available context. It also makes clear why a single instruction can lead to much more model activity than the short prompt suggests.
The $550 estimate covers the model run at API rates. A budget for taking any such prototype further could also include human playtesting, later iterations, support, and release work. Anthropic announced Claude Fable 5.2 on November 9, the same day its engineer shared the build. Treat Saltline as a documented launch-day demonstration from someone at the company that made the model.
Where did the human work come in?
Rieseberg chose the model, set the task, and specified the scope. He says he gave the model no feedback during the initial run. TweakTown reports that he then tweaked some models and gameplay in a later version. The two playable links make it possible to inspect those as distinct milestones.
What to borrow for a first game
Borrow the small scope first. Give a coding agent one room, one player objective, and one reason the objective might fail. Make it possible to restart without reloading the editor. Play the result yourself, and give it to someone who hasn’t read the brief. If that person can tell what to do and finishes or restarts without help, you have something worth polishing.
Our Claude and Godot prototype breakdown applies the same principle to a different kind of test.