Fable 5.2 Built a $550 Game on 165x More Input Than Output

Felix Rieseberg, whose X profile says he builds things at Anthropic, posted on 9 November that he tested Claude Fable 5.2 “with one-shot prompts without human feedback” and that the resulting game cost “~$550 at API pricing, ~2.7M output and 445M input tokens, heavily cached”. The run “Took 8 hours”, he wrote. That is about 165 times more input than output, by our arithmetic.

At Fable 5.2’s list prices on Anthropic’s pricing page, the 445 million input tokens would cost $3,560 if all were billed as fresh input, 6.5 times his whole bill, by our arithmetic. The post had about 2.7 million views when we checked on 10 November.

Business Pill · Tokens Explained

An evergreen Business Engineer explainer on tokens. It teaches the general idea only and does not discuss this game.

What the Post Says

Rieseberg wrote that “Everything in this game is from scratch: models, sounds, textures” and linked a playable version on claude.ai. He did not publish the prompt, a cost breakdown or how much of the input was served from cache, beyond saying it was “heavily cached”.

TechPowerUp reported the game on 10 November under the headline “Anthropic’s Fable 5.2 One-Shots a Playable FPS Game in Just 8 Hours”.

The Same Tokens, Three Bills

Anthropic lists Fable 5.2 at $8 per million input tokens, $40 per million output tokens and $0.20 per million cache-read tokens, with cache writes at $10. We reported the same $8 and $40 list in our Fable 5.2 pricing piece.

The 2.7 million output tokens come to $108 at $40, which is 19.6% of $550, by our arithmetic. If all 445 million input tokens were cache reads at $0.20, input would cost $89 and the run $197. Billed entirely as fresh input at $8, the run would cost $3,668. His reported $550 sits between the two, leaving about $442 for input and cache writes, by our arithmetic.

Over eight hours, $550 is about $68.75 an hour, by our arithmetic.

How Others Read It

Aakash Gupta wrote on 10 November that “the AI read 165 times more than it wrote” and said of his own tests that one Opus run “cost $1.99” while “The answer itself, the part you’d actually read, was 13 cents”. He added: “I’d bet most of that $550 was reading too”. Those are his tests and his estimate.

The Structural Read

As we read it, the bill for this kind of agentic run is set mostly by how much the model reads and how much of that reading is cached, not by how much it writes: output accounts for about $108 of the $550 at list prices, by our arithmetic. Rieseberg did not publish the cache split, so the exact mix is not established.

Three Implications

FOR BUILDERS At Fable 5.2 list prices, cache reads cost $0.20 per million tokens against $8 for fresh input, per Anthropic’s pricing page, a 40-fold difference by our arithmetic.

FOR BUDGETS Priced at list, the same token counts range from $197 with all input cached to $3,668 with none cached, by our arithmetic.

FOR THE PRICING DEBATE Output was 2.7 million tokens, about $108 at $40 per million, or 19.6% of the reported $550, by our arithmetic.

What Is Not Established

How the 445 million input tokens split between fresh input, cache writes and cache reads is not established; Rieseberg did not publish it. Whether the ~$550 reflects list prices exactly or rounding is not established. Whether the game would pass any external quality test is not established, and nobody has independently replicated the run. The priced scenarios use Anthropic’s list prices as read on 10 November.

The Bottom Line

An Anthropic builder reported a Fable 5.2 game made from one-shot prompts for about $550, using about 445 million input tokens and 2.7 million output tokens over eight hours. At Fable 5.2 list prices, output accounts for about $108 of that, and the same input billed as fresh tokens would cost $3,560, by our arithmetic. He described the run as heavily cached.