Token Doubling Effect

• Entity ID: ent-20260410-d6e0425480cd
• Type: concept
• Scope: shared
• Status: active

Description

The hidden cost phenomenon where extractMemories fires a separate Opus API call after every turn, transmitting the entire conversation with different tool definitions and spawning a second independent cache chain. A 20-turn session with 650K context consumes ~26M tokens instead of ~13M. Invisible to users on flat-rate Max plan; devastating on API billing.

Key claims

• 120K token-stop pattern derived from 200K*83% compaction threshold

Relations

• Two-Version Output Efficiency Directive --[caused]--> Token Doubling Effect

Sources

none