Unverified50% confidenceEventExact time
The case in GitHub issue #87303 shows: a session contained 521 usage records but corresponded to only 323 distinct message IDs; directly summing gave cache_creation_input_tokens of 1,075,420, while deduplicating by ID gave 560,751, an overestimate of 1.92 times; the cache-read field was overestimated by 1.59 times.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/19/2026
First Seen
Valid until: 1/17/2027
Sources
Related Entities
Related Claims
UnverifiedGitHub issue #7714 calculated that injecting the same state into the context window consumes 5,764 tokens, while the semantic content itself requires only 3,587 tokens, meaning a 37.8% token overhead per LLM call.77% similarUnverifiedGitHub issue #64880 (still open) reports that under high server load, six independent sequential responses returned the same message.id; the reporter later re-checked during a normal-load window and found no instance of the same ID appearing under multiple requestIds.74% similarUnverifiedA server can return free-text instructions when a client connects; GitHub's default toolset instructions are 377 tokens, its full toolset's are 1,559 tokens, and Terraform's are 924 tokens.72% similarUnverifiedOf GitHub's default-configuration 11,207 tokens of definitions, the description portion accounts for 1,368 tokens (12%) and the input schema for 9,287 tokens (83%).70% similarUnverifiedGitHub issue #66947 reports that each compaction or continuation appears to rewrite earlier records into the same transcript, using the same uuid and timestamp, with one message appearing five times in one file; Anthropic has not confirmed this issue.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1067071API
curl https://kongchang.com/api/v1/knowledge/claims/1067071MCP
get_claim(id=1067071)