I have been trying to run some very simple API°hard word calls using Claude°hard word to operate°hard word the new Barsukas°hard word HTTP°hard word API°hard word.
The flow°hard word is a 20-word prompt°hard word, reading 3-4 files°hard word from disk°hard word, and making a 4-line Python°hard word script°hard word to call the API°hard word.
This costs°hard word 22 cents°hard word. ⚙️ which feels like too much
Why so much? There is something like 8 KB°hard word of system prompt°hard word and 14 KB°hard word of system tools. All of which should be cached°hard word/built-in°hard word to the model, but is instead just charged°hard word to the user°hard word. And several rounds of "cached°hard word reads" as the tool pauses°hard word for inputs°hard word.
Fortunately°hard word, the $20/month subscription°hard word gives something like $300/month of credits. When the cost°hard word is inflated°hard word by a factor°hard word of 5, it helps that the currency°hard word is also inflated°hard word.
----
On the other hand, the sheer°hard word volume of struggles°hard word Claude°hard word Opus°hard word is having with this task (simple directions like "only read the api°hard word/ dir°hard word" get ignored°hard word about 1/4 of the time) does make me want to look for a different harness°hard word, which can pivot°hard word between model providers°hard word. ⚙️ it's not as bad as it was six weeks ago, but it is still very disappointing°hard word for the "recommended°hard word/expensive°hard word model" that it can't do this well.