← All notes

Summary

Tool-schema filtering saves a fixed block of tokens every turn and is the only lever that is unambiguously and reproducibly positive; on a typical turn it removes about 21K to 57K tokens. Content compression, by contrast, saves only about 2% of the cache-priced prefix per turn. The 86.5% SWE-bench quality retained at a 25.7% compression rate is orthogonal to multi-turn agent cost. [S1]

Claims

• id: C1 claim: 'Tool-schema filtering is described as a fixed per-turn token saving.' source: S1 evidence: "Tool-schema filtering saves a fixed block of tokens every turn"
• id: C2 claim: 'They report about 21K to 57K tokens removed on a typical turn.' source: S1 evidence: "on a typical turn it removes"
• id: C3 claim: 'They report 86.5% SWE-bench quality retained at a 25.7% compression rate.' source: S1 evidence: "86.5% SWE-bench quality retained at a 25.7% compression rate"

Verification

• Code available: Authors say the compression model and evaluation harness are Apache 2.0 at https://github.com/Paritok-official/paritok-4b-v1 (HTTP 200 on 2026-11-11). Not run. The 86.5% figure is described as reported separately for Paritok-4B.
• Reproduced: not reproduced (2026-11-11). Code not run.
• Related work: Live note ao-cost-compression-accounting. Beyond Token Savings in this batch looks at success and latency, not only tokens.
• Limits: The study instruments one production gateway (Paritok) on Claude Code and Codex sessions. Savings vary with turn count.
• 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 19 Aug 2026 (v1). Comments: 10 pages, 1 figure, 4 tables. Sources: [S1] https://arxiv.org/abs/2609.22114, v1 (19 Aug 2026). checked_on: 2026-11-11.

When to apply

Use when asking whether a compression gateway actually cuts a coding agent's token bill, and which part of the stack moves the cost.

Facts

• id: F1 claim: 'Tool-schema filtering is reported as the only unambiguously positive lever, removing about 21K to 57K tokens on a typical turn.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F2 claim: 'Content compression is reported to save only about 2% of the cache-priced prefix per turn.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F3 claim: 'A single-shot 86.5% SWE-bench quality score at a 25.7% compression rate is described as orthogonal to multi-turn agent cost.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0

Sources

• [S1] arXiv 2609.22114, https://arxiv.org/abs/2609.22114. License: CC BY 4.0. v1 (19 Aug 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_s56b10fe,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{An Empirical Cost Attribution of Context-Compression Gateways in Multi-Turn Coding Agents}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/2609-22114-compression-gateway-cost/},
  note = {Permanent ID iil:note:s56b10fe. 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/2609-22114-compression-gateway-cost.md.