Unverified50% confidenceFactExact time
Anthropic measures Claude usage in units of message count or the underlying token consumption, and when the context window is heavily occupied, a single conversation can consume several times the normal quota.
1
Sources
50%
Confidence
Long-term
Relevance
9/16/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedThe computational load of the attention mechanism is proportional to the square of the sequence length, meaning doubling the context increases computation fourfold.66% similarUnverifiedThe token consumption per response in reasoning mode increases multiplicatively, with deep reasoning consuming thousands or even tens of thousands of tokens.61% similarUnverifiedInvoking multiple experts simultaneously causes token consumption to multiply, with total token consumption growing linearly or even super-linearly with the number of experts60% similarUnverifiedThe computational complexity of the standard self-attention mechanism is proportional to the square of the context length (O(n²)), meaning extending the context from 100K to 1M theoretically increases computation by 100 times.60% similarUnverified选择性重计算由Korthikanti等人在2022年的Megatron-LM论文中系统研究，仅对注意力激活值做重计算可在额外计算开销不超过10%的条件下接近全量重计算的显存节省60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/924623API
curl https://kongchang.com/api/v1/knowledge/claims/924623MCP
get_claim(id=924623)