Claude Fable 5.1 API pricing and cost calculator

How this is calculated

Begin with the calculator's starting request of 100,000 input tokens and 5,000 output tokens. Uncached, that costs $1.25: $1 for input and $0.25 for output. With half the input read from cache it costs $0.7625, and with all of it read from cache it costs $0.275. At 100 requests a month the starting chart total is $76.25. An agent session is a better test of the model than one request. Picture 40 turns, each sending about 120,000 input tokens and returning 4,000 output tokens, with most of the prompt read from cache because each turn only adds a little to what came before. That session costs $13.88. Without caching, the same 40 turns would cost $56, which shows why a stable prefix matters so much here. Cache writes aren't in either figure. A five minute write costs $12.50 per million tokens, and a one hour write costs $20 per million, twice the input rate. In a long session the prefix grows every turn, so a share of each turn's input is a fresh write and not a read. Knowledge work leans the other way. Reading a 600,000 token document set and writing an 8,000 token memo costs $6.40 with no cache hits, and a follow up question that reuses most of the set from cache costs $1.135. The context window is 1,000,000 tokens, so that fits in one request. The maximum output is 128,000 tokens. A reply that long, with a 20,000 token prompt, costs $6.60 on its own, so set a sensible output cap on automated runs. Input can be text, images or files.

Verdict

Price Fable 5.1 by the finished task, not by the call. Log every turn of a few real sessions, including cache read tokens, cache write tokens, output tokens and retries, then price those logs in the calculator. Check how many turns it takes to reach an accepted result, because a model that needs fewer turns can cost less per task even at a higher rate per token. Work that can wait, such as a nightly code review or a bulk document pass, can use batch prices and skip the interactive rates entirely. Keep prompts tight, put stable instructions first so they cache, and cap output on unattended runs. A smaller model can take the simple steps while this one handles the hard ones.