Build Detail

The first run on this page using OpenAI’s cheap Luna tier: gpt-5.6-luna at xhigh reasoning, driven by Codex CLI 0.147.0 in multi-agent v2 team mode. The brief is the same one every build here gets — implement the whole shop from specs/* in one go, test with Pest, drive the acceptance criteria through the Playwright MCP — with one line that visibly shaped the run: “You must use sub-agents!” At $0.20/M input and $1.20/M output, Luna is roughly a twelfth the token price of the Sol tier used in #15, which is the whole reason this build is interesting.

It did use sub-agents — 23 of them, each auto-named after a scientist or philosopher — but not to write the shop. Reading their briefs, 16 of the 23 are explicitly read-only auditors (“Perform a read-only audit… Do not edit files”), 3 are narrow post-audit fixers (“Fix only API/security transport issues”), and only 4 were ever authorised to implement anything — Einstein on the tenancy foundation, Singer on checkout, Banach on seeders, Copernicus on admin UX, each fenced into a disjoint write scope so parallel writers could not collide. The orchestrator kept the bulk of the work for itself: it made 367 of the 457 apply_patch calls in the entire run, against 90 from all sub-agents combined. The delegation pattern is essentially fan-out for verification, not for labour — wave after wave of independent auditors re-reading the same specs against the same tree, then a small number of tightly-scoped fixers sent in to close what they found. It is a defensible reading of the instruction, and it is also why the sub-agent fleet costs so little: auditors read, and reads are 97.4% cache hits.

The economics are the headline. $9.85 for the entire run — orchestrator and all 23 sub-agents — in 4h 22m, with 340.9M input tokens of which 332.0M were cache reads. Against the builds that score comparably, that is 4× cheaper than #16 Cursor ($41), 12× cheaper than #15 Codex Sol Ultra ($120, for 2.1 more points), 19× cheaper than #17 Kimi K3 ($190), and 43× cheaper than #14 Fable 5 ($423). And yet the delivered repository does not boot. Commit fb83f734 added config/sanctum.php, which calls Sanctum::currentApplicationUrlWithPort(), and put the HasApiTokens trait on App\Models\User — but never added laravel/sanctum to composer.json or the lockfile. At no point in the branch’s history is it a declared dependency. On the build machine this stayed invisible: a stale vendor/laravel/sanctum left over from earlier work in that directory satisfied the imports, so artisan route:list kept working and the run’s own Pest suite finished 95 tests green. On a clean checkout, in CI, or on a deploy, Composer follows the lockfile, the package is simply absent, and — because Laravel eagerly loads every file in config/ on boot — every route on both the storefront and the admin dies with a fatal error before routing even starts. The 94% weighted (132 of 143) reported here is therefore measured after installing that one missing package by hand; as committed, a fresh clone scores nothing at all, because nothing loads. The same shape of defect recurs twice more inside the app itself, where LazyLoadingViolationException on un-eager-loaded relations takes out every product variant selector and, depending on cart contents, checkout too.

Net: the best price/quality ratio on the page by a wide margin, and a precise lesson about the limits of review. Sixteen independent auditors re-read this tree for pocket change and every one of them signed off on a repository that cannot be installed — not because they failed to run it, but because they all ran it in the one environment where a leftover vendor/ directory hid the omission. Nobody checked that the declared dependencies still matched the code’s imports, which is precisely the check a human team hands to CI rather than to reviewers. Cheap verification bought a great deal of reading and not one clean-room build.