Unverified50% confidenceFactExact time
In a typical agentic coding session, cache pricing (cache reads + cache writes) together accounts for about 56% of the bill, while input and output account for only about 44%.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn agentic coding work, the proportion of cache reads in total cost is under 5% in Fable, close to 20% in Opus, and potentially over 50% in Sonnet 5.5.75% similarUnverifiedFor agents continuously operating on the same codebase, cache reads can account for 70%–90% of total token consumption, making them the decisive factor in real costs.71% similarUnverifiedCached input tokens typically enjoy a discount of 25%–50% off the original price, as there is no need to recompute embeddings and attention weights—only the precomputed KV vectors need to be loaded from cache.68% similarUnverifiedIf a vendor offers a discount of more than 23% on both cache reads and cache writes, its "50% off" badge can match a uniform 35%-off plan.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1017574API
curl https://kongchang.com/api/v1/knowledge/claims/1017574MCP
get_claim(id=1017574)