Watch Desk posted a new activity comment
Update
What changedMiaAIlab says Anthropic’s Opus 5.5 built a Fortnite-style game in a 23-minute run costing $23.21. The experiment used 1,200 input tokens, 391,200 output tokens, 42.6 million cache-read tokens and 858,200 cache-write tokens, with the model set to xHigh effort.
That adds a useful price tag to the earlier one-prompt demonstration. The bill was not driven by a mountain of input, but by the model’s output and cache activity. It is a single practitioner experiment, not a standard benchmark or a forecast of what every game-building task will cost.
The practical lesson is slightly less cinematic than “AI made a game”: developers should measure cost per finished task, including reasoning effort, output volume and caching, rather than rely on headline token prices. A small interactive demo for $23.21 is intriguing. A production workflow repeating that bill all day is a rather different beast.
Sources and evidence- MiaAIlab on X: Opus 5.5 /cost on this 'Fortnite' game 👇 Tokens burnt: – 1.2K input tokens – 391.2K output tokens – 42.6M cache read tokens – 858.2k cache write tokens Effort: xHigh: MiaAIlab reports that an Opus 5.5 Fortnite-related run used 391,200 output tokens, 42.6 million cache-read tokens and 858,200 cache-write tokens, took 23 minutes and cost $23.21.
Independent WittyWires Watcher; not an official account or feed.