{"slug": "i-estimate-reading-code-costs-2-1x-more-than-writing-it", "title": "I estimate reading code costs 2.1x more than writing it", "summary": "A developer estimates that human code review costs $0.243 per changed line, 2.1 times the $0.114 model spend for an AI agent to produce an accepted changed line, based on a SmartBear/Cisco study, U.S. Bureau of Labor Statistics wage data, and an OpenHands GPT-5.5 SWE-bench Verified run. The analysis, which excludes an outlier patch, finds the agent resolved 390 of 499 tasks at a total model cost of $577.48 and 128,279 tokens per accepted changed line.", "body_md": "I hate reading agent-generated code and it's not a good use of my time, and I wanted a staistic to quantify by how much.\n\nMy current estimate is $0.243 for a human to review one changed line and $0.114 in model spend for an agent to produce one benchmark-accepted changed line. That makes human review about 2.1 times as expensive under these assumptions.\n\nThe agent calculation uses one public GPT-5.5 OpenHands run from May 2026. Its token categories, final patches, per-task costs, and evaluation outcomes come from the same 500 attempts. I then reprice those measured tokens at GPT-5.6 Sol's current rates.\n\n*Human review: $0.24 per changed line*\n\nA SmartBear/Cisco study analyzed more than 2,500 reviews covering 3.2 million lines of code. It found that defect detection deteriorated as inspection rates rose through the 300-500 LOC/hour range. I used 400 LOC/hour as a midpoint.\n\nThe US Bureau of Labor Statistics reports a median software-developer wage of $65.38/hour for May 2025. BLS employer-compensation data for March 2026 says benefits represented 32.7% of total compensation for private-industry professional and related occupations.\n\nTreating $65.38 as the wage component gives:\n\n```\n  loaded hourly compensation  = $65.38 / (1 - 0.327) = $97.15\n  review cost per LOC         = $97.15 / 400         = $0.243\n```\n\nOpenHands publishes its complete GPT-5.5 SWE-bench Verified run. It used OpenHands v1.18.1 with GPT-5.5 at high reasoning effort and attempted all 500 benchmark tasks.\n\nI counted changed LOC as additions plus deletions in each unified diff, excluding the +++ and --- file-header lines. \"Accepted\" means the SWE-bench evaluator marked the task resolved.\n\nThe run resolved 391 of 500 tasks. One successful result, psf__requests-1142, contained 16,180 changed lines. The next five largest successful patches contained 136, 129, 86, 81, and 79 lines.\n\nThat single patch accounts for 76% of all successful changed LOC. Including it would reduce the aggregate cost from $0.114 to $0.027 per LOC, which feels wrong. I therefore treat it as an outlier and removed that entire task from both cost and LOC totals. I report both results below.\n\nAfter removing it, the dataset contains:\n\n```\n  attempted tasks                    499\n  resolved tasks                     390\n  resolve rate                     78.2%\n  accepted changed LOC             5,077\n```\n\nThe remaining 499 attempts consumed:\n\n```\n  cumulative tokens          651,273,155\n```\n\nGPT-5.6 Sol's promotional prices through November 21, 2026 are $4/M fresh input, $0.40/M cache reads, and $20/M output. Applying those rates:\n\n```\n  total spend over 499 attempts              $577.48\n```\n\nAll failed attempts remain in the $577.48 numerator. Only code from the 390 resolved tasks enters the denominator:\n\n```\n  cost per accepted changed LOC = $577.48 / 5,077\n                                = $0.1137\n\n  tokens per accepted changed LOC = 651.27M / 5,077\n                                  = 128,279\n```\n\nSource: https://static1.smartbear.co/support/media/resources/cc/book/code-review-cisco-case-study.pdf\n\nSources: https://www.bls.gov/news.release/ocwage.t01.htm https://www.bls.gov/news.release/ecec.t04.htm\n\nRun metadata and archive link: https://github.com/OpenHands/openhands-index-results/tree/main/results/GPT-5.5\n\nSWE-bench Verified source data: https://huggingface.co/datasets/SWE-bench/SWE-bench_Verified\n\nComments URL: [https://news.ycombinator.com/item?id=49456758](https://news.ycombinator.com/item?id=49456758)\n\nPoints: 3\n\n# Comments: 0", "url": "https://wpnews.pro/news/i-estimate-reading-code-costs-2-1x-more-than-writing-it", "canonical_source": "https://news.ycombinator.com/item?id=49456758", "published_at": "2026-08-26 22:25:06+00:00", "updated_at": "2026-08-26 22:48:58.781849+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-tools", "developer-tools"], "entities": ["OpenHands", "GPT-5.5", "GPT-5.6 Sol", "SmartBear", "Cisco", "U.S. Bureau of Labor Statistics", "SWE-bench Verified"], "alternates": {"html": "https://wpnews.pro/news/i-estimate-reading-code-costs-2-1x-more-than-writing-it", "markdown": "https://wpnews.pro/news/i-estimate-reading-code-costs-2-1x-more-than-writing-it.md", "text": "https://wpnews.pro/news/i-estimate-reading-code-costs-2-1x-more-than-writing-it.txt", "jsonld": "https://wpnews.pro/news/i-estimate-reading-code-costs-2-1x-more-than-writing-it.jsonld"}}