← All notes

Summary

This study argues that fewer tokens after compression need not mean a faster or cheaper agent run. Across nearly 35,000 agent runs, policies using roughly one-third as many tokens can take 20-80% longer than the uncompressed agent on Terminal-Bench with Qwen. A cited 13 million GitHub Copilot-session trace found that sessions requiring compaction accounted for 44.2% of tokens served, and the median compaction event removed 72.8% of existing context. [S1]

Claims

• id: C1 claim: 'The study covers nearly 35,000 agent runs.' source: S1 evidence: "Across nearly 35,000 agent runs, we measure task success, token use, end-to-end latency"
• id: C2 claim: 'Fewer tokens can still mean a longer run: one-third the tokens, 20-80% longer.' source: S1 evidence: "policies using roughly one-third as many tokens can take 20–80% longer than the uncompressed agent"
• id: C3 claim: 'They report a 44.2% share of tokens in sessions that required compaction, citing a Copilot trace.' source: S1 evidence: "sessions requiring compaction account for 44.2% of all tokens served"

Verification

• Code available: No official repository URL was found on the abs page or in the HTML text searched on 2026-11-11.
• Reproduced: not reproduced (2026-11-11). Code not run.
• Related work: The gateway-cost paper in this batch attributes token bills. This paper adds success, latency, and estimated cost.
• Limits: The 44.2% and 72.8% figures are cited from Liu et al., 2026, not newly measured in the 35,000-run study. No official code URL was found.
• License: CC BY 4.0 on the abs page (2026-11-11; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 26 Sep 2026 (v1). Sources: [S1] https://arxiv.org/abs/2609.32961, v1 (26 Sep 2026). checked_on: 2026-11-11.

When to apply

Use when asking whether turning on compaction will actually save time or money, not only tokens.

Facts

• id: F1 claim: 'The authors measure success, tokens, latency, and estimated cost across nearly 35,000 agent runs.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F2 claim: 'On Terminal-Bench with Qwen, policies using roughly one-third as many tokens can take 20-80% longer than the uncompressed agent.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F3 claim: 'They cite a 13 million-session Copilot trace in which sessions requiring compaction accounted for 44.2% of tokens served, and the median event removed 72.8% of context.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0

Sources

• [S1] arXiv 2609.32961, https://arxiv.org/abs/2609.32961. License: CC BY 4.0. v1 (26 Sep 2026). Accessed 2026-11-11.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.

@misc{iil_note_63p6rhv5,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{Beyond Token Savings: A Systematic Study of Context Compression in LLM Agents}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/2609-32961-beyond-token-savings/},
  note = {Permanent ID iil:note:63p6rhv5. License: CC-BY-4.0.},
}

Markdown

Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.

For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/2609-32961-beyond-token-savings.md.