The quiet release

PixelScale is out, but barely. This is a silent release: shared with a handful of people, no announcement, while the prices get stress-tested in the open.

Will Bilton · Bilton Projects · 2 June 2026

Test build · prices are a starter beta

If you are reading this, you are early. Nothing here is final. The point of a quiet release is to let real people poke at the numbers before anyone shouts about it. So poke at them. If a price looks wrong, it probably needs fixing, and there is a button for that at the bottom.

Why this exists

TokenScale started because "$5 per million tokens" means nothing to a human. A whole novel for six cents lands. So I built a tool that translates token prices into things you recognise.

Creative AI has the same problem, only worse. Nobody prices an image the same way. Some charge per image, some per credit, some per second of video, some bury it inside a monthly plan. Credits mean different things on every platform. PixelScale does for pictures and video what TokenScale did for words: it turns the made-up units back into real projects and real money.

"A 32-page picture book for 64 cents lands. So does the part nobody mentions: the same short video that costs a couple of dollars one way and ten times that another."

The thing that surprised even me

Here is the moment that made me sure this was worth building. I remembered making loads of video on a $20 a month plan last year. Then I checked what a 15-second reel costs today and was genuinely shocked.

It turns out I was not misremembering. The $20 plan used to run a lighter model with looser limits, so I could iterate all day. The new video models are far heavier, so they sit behind credit gates and rolling compute windows. The twist: some raw per-second prices fell while the plan felt tighter. The model could get cheaper while the subscription experience got stingier. That gap, between what you feel you pay and what you actually consume, is the whole reason PixelScale exists. There is a section about it on the tool, called The catch.

What is real, and what is a guess

I want to be straight about this, because creative pricing is slippery. The per-unit prices are sourced from official provider docs where possible, and each row is labelled Official or Estimate. The project totals are estimates, built from assumptions you set yourself: how many finals you need, how many tries each one takes, how many edits. Move the sliders and the number is yours, not mine.

Two honest caveats. Subscription tools like Midjourney and Firefly are labelled as estimates, because a flat monthly fee has no clean per-image price. And the quality tiers scale prices by a representative multiplier, not exact per-provider resolution pricing. When in doubt, the tool shows its working and links the source.

The July audit — what we got wrong, on the record

On 1 July 2026 every row on the board was re-verified against its official source in one sweep, by Claude Fable 5 — the same audit TokenScale ran the same day. The good news first: all six officially-priced models checked out exactly. Imagen 4 Fast, FLUX 1.1 Pro, Ideogram 3.0, Gen-4 Turbo, Veo 3.1 Fast, Sora 2 — every nightly-tracked price was right.

The estimates were less tidy, and you deserve the specifics. ElevenLabs had cut its API price and we hadn't noticed — our per-character estimate was running about 1.8× high; it's now $0.0001 per character, matching their pay-as-you-go rate. Our Suno estimate didn't derive from any current plan — re-derived from today's Pro plan it lands at roughly $0.02–$0.04 per song, not $0.12; we now show the conservative end. And our Runway source link pointed at a docs page Runway had quietly left to rot with stale prices on it — repointed to the live one. Fable found these, Fable fixed these.

Also on the watch list, for honesty's sake: Google retires Imagen 4 on 17 August (its successor line is Gemini's native image generation), FLUX 1.1 Pro is now officially "previous generation" behind FLUX.2, and Runway has shipped Gen-4.5 at $0.12/second. Prices here are still right — but the ground is moving, which is exactly why the six official rows get re-verified every night.

How it was built

Same way as TokenScale. On a phone, by someone who is not a developer, using the AI it was built to compare. One HTML file. No framework. No build step. No server. No analytics script. It runs entirely in your browser, and the only thing it remembers is your palette choice.

The family

PixelScale is the visual sister of TokenScale. Same warm palette, same gauge mark, same honest, manually checked, no sign-up philosophy. Token for words. Pixel for pictures. Both under Bilton Projects.

Thanks for being early.