Who do you actually work with?

Expert-led businesses: practices, consultancies, advisory and specialist firms where the expertise is the product. Established rather than idea-stage, usually five to fifty people. Big enough that growth has stalled in a way costing real money, small enough that a full-time product or growth lead is overkill.

Do we need AI at all?

Often not yet. Most of the value in the first ninety days comes from connecting what you already have and measuring it. AI earns its place where it removes hours from your most expensive people. I will tell you which it is.

How is this different from hiring an agency?

Agencies sell hours and deliverables. I sell judgement and outcomes, and I will tell you not to build something. Most agencies will not, because the build is the product they are selling.

What if the map says build nothing?

Then that is the answer, and you have saved yourself a build. It happens. You keep the map either way.

Who owns what we build?

You do. The code, the content, the data, the repositories. Nothing is rented and nothing is locked in.

Three near-identical wooden chairs standing in a row in a sunlit corner

← All writing

Rebuilding in Public · 4 min

Three AI models, one brief: what each build actually cost

Fable 5.1, GPT-5.6 Sol and Gemini 3.8 Flash each built the same site from the same CV. The runs cost $23.57, $10.20 and $1.48. The harness still decides.

Published 3 September 2026 · Updated 12 September 2026

Three frontier models landed in one week, so I gave all three the same job: here is my CV, build my portfolio website. One pass each, no follow-ups, no fixes, each inside its own coding harness.

The headline is not the ranking. The strongest build cost the most and took the longest to think. The weakest was finished in seventeen minutes for a dollar fifty. And every one of them did genuinely complex work (structure, copy, styling, a working site) for less than a round of coffees. The price of complex work is the story.

What launched, and in what

Model	Lab	Run inside
Fable 5.1	Anthropic	Claude Code, on high
GPT-5.6 Sol	OpenAI	Codex, on high
Gemini 3.8 Flash	Google	Antigravity

Meta's Muse Spark 1.3, its most powerful yet, shipped in the same week but did not go into this test.

What each run cost

What each run cost, to scale

Fable 5.1 in Claude Code

Dollars 23.57

GPT-5.6 Sol in Codex

Dollars 10.2

Gemini 3.8 Flash in Antigravity

Dollars 1.48

Sixteen times the bill for the same brief. The table adds time and tokens:

Run	Cost	Time	Effective tokens
Fable 5.1 in Claude Code	$23.57	47 min	~11M
GPT-5.6 Sol in Codex	$10.20	29 min	~2.6M
Gemini 3.8 Flash in Antigravity	$1.48	17 min	~1.8M

Costs as metered by each tool where shown, otherwise at list API prices. Effective tokens discount cached re-reads.

Which model built the best website?

On a one-shot prompt for this specific job, web development, Fable 5.1 is the clear winner. It is also by far the most expensive, the hungriest on tokens, and the slowest to deliver.

GPT-5.6 Sol did better on overall interface and experience. Gemini 3.8 Flash came back weakest of the three, though it is delivering roughly what last generation's frontier models gave you, about Opus 4.6 or 4.8 level in my experience, at a fifteenth of the price.

All three builds are live:

The Gemini 3.8 Flash build, a dark terminal-style hero
The GPT-5.6 Sol build, a grid-paper hero
The Fable 5.1 build, a blueprint-style hero

So what does a model launch mean for your business?

Less than the launch posts suggest, and this is the part worth taking away.

My test measured pure foundational capability: one prompt, no help, no corrections. That is a fair way to compare engines and a terrible way to run a business. You do not need to build one-prompt websites, and that was never the point.

The point is what these engines do inside a harness built for your business: your instructions, your context, your checks, your sign-off on everything that leaves. Build that, and a week like this one made every well-built harness better without anyone lifting a finger.

New models rent well. Process compounds.

Also worth knowing from that week

Video, generated faster than you can watch it. A retrained video model now makes fifteen seconds of footage in about nine seconds. Press play and the film does not exist yet; it is being made ahead of you as you watch. Already being used for genuinely interesting things, like prompting the lesson you want to learn and having it taught back to you as video. Also for an ocean of slop.

The mystery model was not Google's. A free, unnamed model called 0xAlpha got so good so fast that the internet decided only Google could afford to run it, and more than a few Google people spent the week nodding along at the secret next Gemini. Then the mask came off: GLM-5.3-Flash, an open model from Chinese lab Z.ai, free to use commercially, running on Chinese chips. Takes have rarely been deleted faster.

Common questions

Which AI model is best for building websites? On a single unassisted prompt, Fable 5.1 produced the best result in this test. It was also the most expensive at $23.57 and the slowest at 47 minutes.

How much does it cost to have AI build a website? In this test, between $1.48 and $23.57 for one pass at a complete working site. The cheapest run took seventeen minutes.

Should I switch models when a new one launches? Usually not. A launch improves every well-built harness automatically. If switching model is more than a settings change for you, the thing to fix is the harness, not the model.

Is a cheaper model good enough? Often. Gemini 3.8 Flash came back weakest here and still delivered roughly last generation's frontier quality at about a fifteenth of the cost. The right question is what the work is worth, not which model tops a benchmark.

The argument underneath all of this, with the business examples, is in language models reason, code calculates. If you would rather build it than read about it, the free course is open to everyone, no sign-in.

This started life as Build Notes Nº 03.