Fable 5.2 runs the same Ice Cathedral: by API-equivalent estimate 76% more expensive, 75 minutes faster

Same Cathedral of Ice prompt, one run each for the two models. Fable 5.2 delivered $81.10 and 1 hour 56 minutes 52 seconds; Opus 5.5 was $45.97 and 3 hours 11 minutes 26 seconds. Based on this set of API-equivalent estimates, Fable costs about 76% more and saves about 75 minutes of waiting.

Tony's exact words were a huge leap, with the emphasis on the details of the ice, boat, and whale. The comment section was not unanimous. Asmir directly pointed out that this is one sample per model from one run, that $81 vs. $46 is 76% extra spend, and that rerunning the same prompt could also vary a lot. Tony acknowledged that a single run does not prove consistency, but what excited him was the visual detail of this result. That response pulls the debate back from model stability to single-run delivery. Both things matter; the test just didn't answer the former.

Costs also need to be seen clearly. In a reply to Johnny Dangler, Tony explained that the costs in the image are API-equivalent estimates; he actually ran both models on a subscription, and the usage limits were painful. Fraser added the pricing from the release notes: standard is $8/$40 per 1M tokens, fast mode is $16/$80 per 1M tokens, with output speed up to 2.5×. Anthropic's release post still recommends starting most coding tasks with Opus 5.5 and leaving Fable for the hardest open-ended work. This recommendation does not conflict with the single Ice Cathedral test: one looks at peak results, the other at the default choice.

A few more details are worth putting on the table. Lee Wyatt Corp asked whether he would still choose Fable next time; Tony said yes, citing the extra detail and shorter wait. Tom Baker asked whether Opus could reach the same level with a few more prompts; Tony said yes, but the key point is one prompt with no follow-up—iteration adds time and may eat up the cost savings. shobhanai asked whether the speedup came from speculative decoding, token throughput optimization, or a geometric sampling shortcut; Tony said this run cannot tell, because the timing includes scene building, rendering, and checking, and the cause was not isolated. Aayan asked whether the software was cooked; Tony replied that the weekly allowance got cooked first—he never studied programming, and he is excited to see what can be built. Johnny Dangler preferred Opus's scene, especially the orca, and thought neither was worth the price. Tony responded that preferring Opus is reasonable, and that he preferred Fable's water and ice.

So what does this comparison show? Under the same prompt, one run each, and no follow-up iterations, Fable 5.2 produced finer ice and water, at the cost of a higher API-equivalent cost and stability that a single sample cannot prove. It is more like a one-off delivery sample. To decide who to choose for everyday coding, Anthropic's own recommendation, number of reproductions, failure rate, and real-task cost are more persuasive than one video. The comment section questioned it first, and that direction is right.

Original tweet: https://x.com/EnvolDev/status/2119907318729245180