{"slug": "macro-execution-reduced-codex-rounds-by-up-to-84-in-a-60-task-benchmark", "title": "Macro execution reduced Codex rounds by up to 84% in a 60-task benchmark", "summary": "A macro execution approach reduced Codex rounds by up to 84% in a 60-task benchmark, according to Tura maintainer and open-source implementer. Testing Codex CLI 0.144.1 with GPT-5.6-sol on 60 DeepSWE debugging and rewrite sessions, the Macro Direct configuration used 83.5% fewer observed tokens and 84.0% fewer rounds than Codex CLI High, while achieving a 63.3% pass rate versus 60.0%. The Macro + backward reasoning configuration reached an 80.0% pass rate with 2,017 rounds and $221.14 estimated cost.", "body_md": "I maintain Tura and open-sourced this implementation. This is a project benchmark, not an independent product review.\n\nA typical repository task can spend separate model turns on environment inspection, several edits, building, tests/lint, and reading Playwright media. The macro approach defines that deterministic workflow once, runs it below the model boundary, and returns the combined evidence when another reasoning decision is actually required.\n\nI tested Codex CLI 0.144.1 with GPT-5.6-sol on 60 DeepSWE debugging and full-repository rewrite sessions:\n\n| Configuration | Passes | Pass rate | Observed tokens | Rounds | Estimated cost |\n|---|---|---|---|---|---|\n| Macro + backward reasoning | 48/60 | 80.0% | 229,695,477 | 2,017 | $221.138 |\n| Macro Direct | 39/60 | 65.0% | 75,108,167 | 969 | $99.620 |\n| Codex CLI Medium | 38/60 | 63.3% | 333,538,349 | 3,140 | $257.173 |\n| Codex CLI High | 36/60 | 60.0% | 455,742,296 | 6,074 | $327.483 |\n\nAgainst Codex CLI High, Macro Direct used 83.5% fewer observed tokens and 84.0% fewer rounds, with a 5-point higher pass rate. The backward-reasoning configuration spent part of the saved budget on stronger investigation and verification and reached 80.0%.\n\nThis is not blind batching: the macro stops whenever new evidence requires a decision. I also would not generalize the aggregate table to every repository; task mix, failure handling, and macro boundaries matter.\n\nImplementation, scripts, and data:\n\nBenchmark record:\n\nWhere would you draw the boundary between a safe deterministic batch and a step that should always return control to the model?", "url": "https://wpnews.pro/news/macro-execution-reduced-codex-rounds-by-up-to-84-in-a-60-task-benchmark", "canonical_source": "https://discuss.huggingface.co/t/macro-execution-reduced-codex-rounds-by-up-to-84-in-a-60-task-benchmark/178055#post_1", "published_at": "2026-07-21 02:17:45+00:00", "updated_at": "2026-07-21 02:32:55.303307+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "developer-tools"], "entities": ["Codex CLI", "GPT-5.6-sol", "DeepSWE", "Tura"], "alternates": {"html": "https://wpnews.pro/news/macro-execution-reduced-codex-rounds-by-up-to-84-in-a-60-task-benchmark", "markdown": "https://wpnews.pro/news/macro-execution-reduced-codex-rounds-by-up-to-84-in-a-60-task-benchmark.md", "text": "https://wpnews.pro/news/macro-execution-reduced-codex-rounds-by-up-to-84-in-a-60-task-benchmark.txt", "jsonld": "https://wpnews.pro/news/macro-execution-reduced-codex-rounds-by-up-to-84-in-a-60-task-benchmark.jsonld"}}