Unverified70% confidenceTradeoffExact time
Unsloth abandoned the conversation summarization approach because the benefit of summarization was minimal yet added approximately 190 seconds of overhead per compression
2
Sources
70%
Confidence
Medium-term (~90 days)
Relevance
9/6/2026
First Seen
Valid until: 12/5/2026
Sources
Related Entities
Related Claims
Unverified/clear completely wipes the current session and starts from zero, with the lowest cost but loss of all context; /compact has the model compress the existing conversation into a summary that retains key information and then continues, still incurring one call's cost but far less than having the whole original text repeatedly participate in subsequent requests.67% similarUnverifiedAfter a long interruption (cache expiry), compaction is actually cheaper than simply continuing the conversation, because each subsequent turn then reads only about 15k tokens instead of the entire context.65% similarUnverifiedUsing /compact between tasks cost $2.34, 17% more expensive than doing nothing, because compaction requires reading the entire context before compressing it.62% similarUnverifiedAccording to the author's logs, actual input content accounted for only about 0.01% of the bill, with 96% of the cost coming from old conversation history being resent repeatedly.59% similarUnverifiedIn the experiment, "starting a new session for a new task" did not save tokens and instead used about 22% more, because the old task was short and still in the cache, and after /clear discarded the cache the new session had to be rebuilt at full price.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/864747API
curl https://kongchang.com/api/v1/knowledge/claims/864747MCP
get_claim(id=864747)