Plansmith
PlanSmith Benchmark

Retail inventory — the model build-off

We dogfood every PlanSmith planner across multiple coding models. We lock the vertical, then audit what each one builds. The planner is the product — these are the apps it produced. Every one is a working retail stock system. Each build is scored 0–100 against its own run's selected-feature ledger: completeness of what that run actually selected (45), engineering quality (30), and UI/UX (25) — so a feature deferred at selection time never counts against a build.

2026-07-104 models · 8 builds audited
The build-off · 01

How the builds ranked.

01

Harbor Lane

Best Overall
by Composer 2.5
93/100

The strongest build on both features and design.

Best at
All eight retail workflows finished — size-and-colour variants, printable barcode labels, receiving deliveries, returns, stocktakes and stock valuation — on a proper database with real password security, behind the best-looking interface of the set. It even scans your own catalogue for messy data.
Watch for
It arrived nearly empty, so the screens looked bare on first open despite everything working.Fixed since auditthe seed now lands populated: purchase orders across states, receiving history, and count sessions on first load.
Open live demo
02

Stockbay

Best Data Integrity
by Grok 4.5
91/100

The second Grok lane — and the set's cleanest domain model.

Best at
The strictest stock discipline of the set: your on-hand figure can only ever change through receiving, an adjustment, a count or a return, and every one of those is recorded with the before and after numbers and who did it. Fifteen screens with nothing left as a placeholder, plus variants, printable labels, a checked spreadsheet import, supplier returns, exports, and one-tap logins that drop each role straight onto its own workbench.
Watch for
It was audited with only nine products loaded, so every screen looked sparse, and its own testing scaffolding was riddled with errors — though the app's own code was clean.Fixed since auditthe demo now opens on a fully stocked shop — 36 products, purchase orders in every state, delivery history, stocktakes showing variance, and a supplier return — and its testing scaffolding no longer contaminates the app's own checks. These fixes were applied in-house: Grok's agent wasn't available for this round.
Open live demo
03

Northstar Goods

Most Complete
by GPT-5.5
87/100

The consistent all-rounder — every surface, no stubs.

Best at
Every one of its fifteen screens actually built — nothing left as a placeholder — with four roles, tamper-proof sign-in, a dashboard you can click through into the underlying records, and a demo that arrives full of real activity: purchase orders, stocktakes, deliveries and returns from day one.
Watch for
It stores everything in plain files rather than a real database, and shipped with a fallback sign-in key baked into the code.Fixed since auditthe server now refuses to start without a real session secret, cookies are TLS-aware, and demo login is one click.
Open live demo
04

StockKind

Best Polish · Zero Dependencies
by GPT-5.5
86/100

The most presentable handover — and it runs on nothing.

Best at
The best-looking dashboard of the set — a stock-health score, where your money is tied up by category, and a reorder queue ranked by urgency — plus a complete stocktake workflow and one-click demo logins, all running with nothing else installed. It built everything its run asked for: receiving and supplier returns were left out at selection time, not dropped by the model.
Watch for
It stores data in plain files and keeps sign-ins in memory, so a restart logs everyone out. Purchase orders originally stopped at “sent” — our first audit wrongly counted the deliberately deferred receiving step as a gap.Fixed since auditreceiving against POs is now live with partial receipts, status transitions, and movement writes — built beyond its selected scope, verified end-to-end.
Open live demo
05

Command Center

Feature-Rich
by GPT-5.5
81/100

Strong everywhere, exceptional nowhere.

Best at
Built around the scanner — receiving, labels, and a queue for barcodes it did not recognise — with permissions checked on the server that hide cost prices from staff who should not see them, over the most richly connected demo data of the set. Supplier returns and analytics were left out at selection time, so neither counts against it.
Watch for
As delivered there was no way to run it as a single working app in production — the back end could not serve the front end at all.Fixed since auditthe server now serves the built app with a safe history fallback on a single port — verified live.
Open live demo
06

Northline Outfitters

Deepest Operations
by GPT-5.5
77/100

The broadest lifecycle in the set, behind the plainest UI.

Best at
The widest day-to-day coverage of any build — receiving, supplier returns, adjustments with a reason attached, stocktakes, labels, scanning, spreadsheet import, stock valuation and eight exportable reports — on a proper 22-table database, running with nothing else installed. Its run also asked for the broadest scope of the set, and it delivered nearly all of it.
Watch for
It shipped with no sign-in at all — a direct failure of the permissions requirement its own run selected, where roles being nothing more than a menu choice is precisely the stated failure. The landing screen is bare and the typography is the plainest here.Fixed since auditreal sign-in with properly stored passwords is now on by default.
Open live demo
07

Larder

Only Multi-Location
by Gemini 3.1 Pro
62/100

Highest ceiling on paper; took three fix rounds to get there.

Best at
The only build that took on more than one location — two shops plus a warehouse, with stock transfers between them, built even though transfers were not required for this run — alongside stocktakes, purchase orders, receiving, reorder rules and a full record of every stock movement.
Watch for
Barcode labels and scanning were asked for and never built. Signing in is just picking a name from a list with no password — exactly the failure its own permissions requirement describes — and it arrived full of test junk, with no way to load proper demo data and a framework bug that silently broke a filter.Fixed since audita realistic seed, the filter bug fixed, and a redesigned shell with a dedicated persona login page.
Open live demo
08

Harbor Lane Market

Late Finish
by Grok 4.5
35/100

Missed the audit window; the finished build is real.

Best at
A complete single-shop operation — purchase orders through draft, sent and part-delivered, receiving against them, stocktakes that show the difference, supplier returns, labels and reasoned adjustments — plus the most transparent dashboard here: every panel names where its number comes from, and a scope list states what was deliberately left out.
Watch for
At audit the entire operational side was placeholders — only the product catalogue, spreadsheet import and history were real — and you signed in by picking a role rather than with a password.Fixed since auditthe finished build now sits mid-table: real password sign-in, a fully populated operation with purchase orders in every state, and working exports restricted by role. On today's merit it would place around fifth; the rank records what we found during the audit window.
Open live demo

Models tested: GPT-5.5, Composer 2.5, Gemini 3.1 Pro, and Grok 4.5. The four GPT-5.5 entries are the same model across different build lanes — not different model versions, and Grok 4.5 ran two lanes: one that missed the original audit window (scored as found) and one that finished later and was audited on its own merits. Feature scope was chosen per run at selection time, so scores measure each build against what its own run asked for — not against the widest build in the set. This is a PlanSmith dogfood benchmark for specific retail-inventory runs, not a universal ranking of coding models. Scores reflect each build at audit time. Audit findings were fixed before the demos went live — by each model's own agent where available, otherwise in-house — and improvements are marked “fixed since audit.”