Fable 5 vs Fable 5.1 across 22,022 of my own API calls: same per call, 31% more tokens per prompt, 31% cheaper per prompt
Heat trend
The percentage is based on available heat signal, not comment count or independent people.
A user compared Fable 5 and Fable 5.1 across 22,022 API calls over 21 days, finding that Fable 5.1 uses 31% more tokens per prompt than Fable 5. However, Fable 5.1 is 31% cheaper per prompt on average, costing $1.05 compared to Fable 5's $1.52. This is because Fable 5.1 bills cached read tokens at 25% of the price of Fable 5, making it more cost-effective despite higher token usage.
I felt like I've started reaching my weekly capacity much faster, and couldn't cleanly tell why. Until I measured per-prompt instead of per-api-call token costs, which on my corpus appeared to be ~31% higher per prompt than with Fable 5.
https://preview.redd.it/b8mf9wfup6nh1.png?width=3000&format=png&auto=webp&s=a2aeb95e17497cb76181a030715ff23bfa89d186
Right after that I priced the same prompts with API list rates and that changed the perspective. Each Fable 5 prompt averages ~$1.52, and for 5.1 average is at $1.05. Thats because almost all of the extras are cache reads and 5.1 bills only 25% of price per cached read tokens compared to what Fable 5 priced. So even with 31% more tokens per prompt - every prompt is still 31% cheaper in average.
And as my weekly bar keeps filling up visibly faster, I have only three theories left of what could be the reason which I can't measure myself just from transcripts, and none of them is tied to model verbosity:
- The plan's limit math may not be passing the cache-read discounts of Fable 5.1
It tends to grab more context when responding to each prompt. And i guess the proper way moving forward is to put in place more guards onto what context should be collected by model and which avoided within my own prompts.
Official 'Prompting Claude Fable 5.1' warns that 5.1 tends to be issuing one tool call per turn instead of batching them. And the proposed fix is the next light nudge addition to your prompts:
First privately list what you need next; then request every item that doesn't depend on another's result in this one response.
Just added it to my own CLAUDE.md. Still need some time to collect data to see if that helps.
I got my numbers from my own session archive, pond, where I collect all my Claude Code sessions from every machine I run it on. They cover 21 days, 22,022 API calls. I can leave queries in a comment if anyone is interested to run analysis on their own jsonl transcripts.