Plansmith
PlanSmith Benchmark

Fleet management — the model build-off

This is the most tightly controlled comparison we have run. All four models were handed a byte-identical planner package — same source hash, same generator, same sixteen selected features — so every difference in the results is the model, not the brief. Fleet software is easy to fake and hard to build: the live map is the most photographed screen in the category, so it is the one models reach for first, while the yard actually needs to know which van is roadworthy this morning, who is driving it, what is overdue, and what it cost per mile. These scores are about whether the app can answer those questions, whether a driver can be stopped from touching another driver's vehicle, and whether the numbers on the dashboard agree with the records behind them.

2026-08-314 models · 4 builds audited
The build-off · 01

How the builds ranked.

01

FleetOrbit

Best Overall
by Claude Fable 5
98/100

The one to copy. Nothing we tried to break came apart.

Best at
Getting the hard parts right without being asked twice. Roles are genuinely separate, and the highest-sounding role is not a superset — the administrator is refused a screen the fleet manager owns, which is the row most builds get backwards. A driver sees one assignment out of four, and it is their own, so the restriction is doing real work rather than returning an empty list. It writes its own checks by hand — 91 of them, every one matching its own record with nothing missing or extra — and ships with no known security problems in anything it depends on. Started against an empty database it sets itself up properly instead of opening onto a broken dashboard.
Watch for
You cannot re-run its own checks to confirm the 91-out-of-91. They use fixed names and change the data as they go, so the second run trips over the first one's leftovers. Nothing suggests the build is broken — but a buyer cannot independently verify the headline claim, which is the point of shipping the checks at all. It also packages a 4 MB uncommitted database sidecar, and one button on a phone is smaller than a thumb.
Open live demo
02

Roadstead Fleet

Strongest Front Door
by GPT-5.6 Sol
93/100

The most carefully secured build here — and the one you can least easily run on your own machine.

Best at
Keeping strangers out. It is the only build of the four with a real sign-in gate: every staff page sends a visitor without a session to a login screen, and an address that does not exist returns a proper not-found instead of quietly showing the app shell. Sign-in tokens and customer tracking links are both stored scrambled, so reading the database does not hand you a working session or a customer's link. Customer contact details live in their own table with an expiry, a revoke and a retention date. It also went through six separate rounds of its own hostile client testing, and its results file records what each check actually proved rather than just the word 'pass'.
Watch for
It will not start with the command it ships with. The build targets a newer version of its hosting runtime than the version its own dependency list installs, so the start command fails immediately on a clean copy. This never showed up during the build because it was deployed to a host that supplies its own runtime. Separately, it packages 93 MB of development telemetry, and running its own QA overwrites the very file that records its results — which is not in version control.Fixed since auditthe start command was repaired — its hosting tooling was brought up to a version whose runtime matches what the build targets, and a clean checkout now builds and serves on its own. The development telemetry is out of the package, which cut it by 80 per cent, and the QA records are under version control so running the checks can no longer erase the file that proves them. One rough edge remains: the dependency lock file was written on Windows and omits the Linux-only entries, so a strict install refuses on a Linux host and has to fall back.
Open live demo
03

FleetFlow

Best Data Model
by Gemini 3.7 Flash
85/100

The richest product here, and the front door is shut now. Rescored after repair, 72 → 85.

Best at
Actually modelling a fleet. It has twenty tables with real fields — fuel with gallons, odometer and cost per fill, geofences with a radius, parts with reorder levels, purchase orders with vendors — where the build above it has one generic record. It is also the only one that wires features to each other: a failed item on a pre-trip check automatically raises a linked critical defect, takes the vehicle out of service and moves the odometer on, and a technician can turn that defect into a work order and close it out.
Watch for
It still shows every role the same records. A driver signs in and reads all five vehicles, every work order and the whole fuel history rather than their own assignment — the build above it gets this right, from the same planner. Writing is properly separated, so a driver is refused the actions that are not theirs; it is reading that is open. The role buttons also hand out a fleet-manager session to anyone who asks for one, which suits a demo and would not suit a yard.Fixed since auditits first audit found it unpublishable: the front end signed every visitor in as the fleet manager on page load, there was no sign-in screen at all, and its first-run setup accepted anonymous requests — a stranger renamed the company and set its currency to a code that is not a currency. It also wrote a fixed fuel-efficiency figure into the record whenever it could not work one out and displayed it under a heading claiming it was calculated, and its cost-per-mile divided the money spent by the total on every odometer in the fleet, making the headline about a hundred times too small. All four were repaired and re-checked on the live demo: an anonymous request now gets refused on every endpoint including setup, a first fill-up reports no efficiency rather than a made-up one, and cost per mile divides by the miles actually driven, with the lifetime figure kept as its own separately named number.
Open live demo
04

Harborline Fleet

Most Reproducible
by Grok 4.6
78/100

Sound foundations, thin product. Good if you intend to fill the features in yourself.

Best at
Being checkable, which nobody else managed. Its full set of 42 checks re-ran cleanly on our machine from an empty database and reproduced its own result exactly — the only build here you can verify rather than take on trust. It also ships a second pretend company in its own data purely so the separation between customers can be proved, and that separation holds on every route we tried.
Watch for
Thirteen of its sixteen features are the same table with different labels. Fuel has no gallons, costs have no amount, geofences have no radius, parts have no quantity — the amount is typed into the record's name. Send it a fuel reading with gallons and an odometer and it accepts the record and throws the numbers away. Routine maintenance can only ever be recorded once: a second oil change is refused anywhere in the fleet, forever. And the dashboard counts rows that do not appear on the screen it links to.Fixed since auditits first audit also found that three of the four roles could not create anything at all and that the administrator account had no working screens behind it. Both were repaired and the build re-audited: each role now does exactly the work its own written contract assigns it, the administrator can manage users, roles and backups, and a new restriction confines a driver to the vehicle assigned to them. That rework moved it from 65 to 78. The generic-record problem above was not addressed.
Open live demo

Models tested: Claude Fable 5, GPT-5.6 Sol, Grok 4.6 and Gemini 3.7 Flash — all four from the same planner package, so all four numbers are directly comparable. Scores are as found at audit, except where a build was reworked and rescored openly: two were, and both movements are shown (65 → 78 and 72 → 85) with the repaired faults still described rather than deleted. All four are now published as live demos, so the spread is open to inspection rather than asserted. These grade correctness, not visual design.