I was reading through a curated list of 60 real-world Claude Fable 5 cases (each logged with input, process, output, and an evidence tag), and one comparison stopped me cold.
Same Physarum simulation, two stacks:
-
GPT-5.5 on Codex: ~17 min, ~$6
-
Claude Fable 5 on Cursor: 40+ min, $360.55
Same task. Roughly 60x the cost. No benchmark table would ever surface that, because the bill is driven by the model × tool × task interaction, not raw capability.
The other thing I took from the list is the recurring “Relay” pattern people seem to converge on: plan and review with the expensive model, route bulk implementation to cheaper ones (4.8, GPT-5.5, Sonnet). Think expensive, build cheap, review expensive.
Curious what others have seen:
-
Have you hit a cost blowup like this on a real task? What was the model/tool combo?
-
Do you actually run a relay/routing setup, or just pick one model and eat the cost?
-
How do you estimate total cost before committing, given unit price tells you almost nothing?
The full case list is here if useful: