{"slug": "token-reduction-is-not-cost-reduction", "title": "Token Reduction Is Not Cost Reduction", "summary": "A pre-registered study of 2,908 Claude Code runs found that token reduction does not reliably lower billed costs for API-based coding agents, with prompt-cache traffic accounting for about 87% of costs and tool-output compression sometimes increasing costs by 6.8% while harming task success. The authors propose evaluating agents by success-adjusted billed cost rather than token reduction alone.", "body_md": "arXiv:2607.12161v1 Announce Type: new\nAbstract: Context-reduction layers for API-based coding agents, including command-output compressors, retrieval rankers, and payload-optimizing proxies, are usually evaluated by how much text they remove. We ask instead: when does reducing retrieved context or tool output lower the actual billed cost of a coding agent without reducing task success or lengthening its trajectory?\nOur primary evidence is a pre-specified, hash-frozen, paired campaign of 2,908 provider-billed Claude Code runs, of which 2,848 were analyzed, covering 103 tasks, seven repositories, and three models. The campaign compared a baseline with two generations of hook-based compression and an API-boundary proxy, within a broader measured program of roughly 5,500 billed executions.\nThree findings emerge. First, prompt-cache traffic dominated cost composition. Cache creation and reads accounted for about 87% of reconstructed four-component cost, or about 80% of the actual bill, with an 8.7% dollar-weighted residual that retained telemetry could not attribute. On Haiku 4.5, this residual scaled with thinking effort.\nSecond, tool-output reduction did not reliably predict billed-cost reduction. An arm that removed 38% of estimated raw tool-output tokens had 6.8% higher paired cost (95% CI: +2.8% to +11.3%), while per-task reduction was only weakly associated with cost change (Pearson r = 0.15, CI crossing zero).\nThird, compression can harm task completion by removing action-critical evidence. In a small single-shot study on SWE-bench-derived Go tasks, compression reduced patch application from 27/40 to 15/40 by corrupting verbatim edit anchors, and the compressed grounded arm produced fewer solves at higher observed cost per solve.\nWe propose a layered evidence standard centered on success-adjusted billed cost rather than token reduction alone.", "url": "https://wpnews.pro/news/token-reduction-is-not-cost-reduction", "canonical_source": "https://arxiv.org/abs/2607.12161", "published_at": "2026-07-15 04:00:00+00:00", "updated_at": "2026-07-15 04:32:49.157221+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-agents", "ai-infrastructure"], "entities": ["Claude Code", "Haiku 4.5", "SWE-bench"], "alternates": {"html": "https://wpnews.pro/news/token-reduction-is-not-cost-reduction", "markdown": "https://wpnews.pro/news/token-reduction-is-not-cost-reduction.md", "text": "https://wpnews.pro/news/token-reduction-is-not-cost-reduction.txt", "jsonld": "https://wpnews.pro/news/token-reduction-is-not-cost-reduction.jsonld"}}