Unverified60% confidenceFactExact time
In large language model inference, the actual hardware cost of output tokens is far higher than that of input tokens, because the decoding phase is a memory-bandwidth-limited sequential process.
2
Sources
60%
Confidence
Long-term
Relevance
7/9/2026
First Seen
Sources
GPT-5.6 Sol限量预览：分级模型Sol/Tera/Luna与准入机制详解
bilibili开水庆余年6/27/2026
Related Claims
Unverified语言模型生成 token 本质上是内存带宽受限任务而非纯算力受限81% similarUnverifiedLarge language model APIs are generally billed by the total number of input plus output tokens, and re-injecting historical context on every call in long projects causes token consumption to balloon rapidly.77% similarUnverifiedWhen running large language models on consumer-grade hardware, the main bottleneck is usually memory rather than the model itself.74% similarVerifiedThe cost of each large language model call consists of two parts: input tokens (Prompt) and output tokens (Completion).74% similarVerifiedThe computational cost of generating output is typically far higher than that of processing input, and some vendors set separate quotas for input tokens and output tokens.74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/366702API
curl https://kongchang.com/api/v1/knowledge/claims/366702MCP
get_claim(id=366702)