Bilton Projects · Field Notes · Updated August 2026
TokenScale by the Numbers
A tool built on a phone, without a programming degree, using the AI it was designed to compare. That was the launch. What runs it today is a system. Here is every number we could find.
Will BiltonBilton ProjectsMay 19 2026 · Updated August 25 2026
TokenScale exists because "$5 per million tokens" never made intuitive sense. At launch in May 2026, saying The Hobbit cost $0.06 on Gemini Flash-Lite made the price tangible. The tool launched with 16 AI providers and now makes that translation across 22 providers and 66 standard tiers, for free, with no sign-up. What follows is the story of how it got built, and how a phone-built tool became a system, told in the only language the project understands: numbers.
The irony
"TokenScale compares 22 AI providers. Claude helped build the launch. ChatGPT joined through Codex for phase two. The tool now compares both companies behind the AI systems used to build it."
At a glance
TokenScale
AI API pricing in words you know
sticker price$5 / million tokenshard to picture
→
launch example · May 2026The Hobbit = $0.06on Gemini Flash-Lite
48,347lines today 16,000 at launch
22providers today 16 at launch
269numbered builds since Pinball began
66standard tiers checked nightly
~46
TokenScale sessions
Traced from conversation history across this Claude account
316
Days of context
From first token pricing curiosity to launch-ready MVP
48,347
Lines of code
16,000 at launch. The current shipping index is still one HTML file.
22
AI providers
16 at launch. 22 tracked today across 66 standard tiers.
6
Nights of data
Entered by hand. At 8pm. On a phone.
0
Dev environment
No terminal. No npm. No git repository.
$0
Infrastructure cost
Cloudflare free tier. Drag-and-drop deploy.
Prices fall, and the record proves it. Every provider's price-history chart is anchored to a pre-tracking baseline, its previous-generation price. Anthropic's mid tier went from $8/$24 (Claude 2.1, 2023) to $3/$15 (Claude 3 Sonnet, March 2024); Opus dropped from $15/$75 to $5/$25. The nightly record picks up where those baselines leave off, so the dataset shows real prices moving over time, both down and (occasionally, like Gemini) up. That's the long game: an independent dated record no one else is keeping in this form.
Token estimates
Anthropic doesn't expose per-conversation token counts in the standard interface, so these figures are honest estimates, derived from session count, average message density, and the fact that file-heavy launch sessions repeatedly uploaded the then-16,000-line HTML file. The same shipping index is 48,347 lines today. Those sessions ran 60,000–100,000 tokens each, so the estimates are likely conservative.
Estimated user prompts sent
~800–1,200
Estimated input tokens
~12–20 million
Estimated output tokens
~8–15 million
Equivalent pay-per-use API cost
~$150–$300
Methodology: ~46 TokenScale-specific launch sessions out of ~200+ total account sessions. Average 18–25 exchanges each. File-heavy sessions dominated token count because the launch-era 16,000-line HTML file was uploaded repeatedly; today's index is 48,347 lines. Equivalent cost calculated at Claude Sonnet rates ($3/$15 per MTok input/output). All of this was covered by a $20/month Pro subscription.
What makes this unusual
"Every session was on mobile. Most were voice-initiated. The nightly price data is entered manually at 8pm. There's no local dev environment, no git, no build step. The whole product is one HTML file."
— Will Bilton, Bilton Projects
"Maybe you should give people clear advice about which AI would be best for them… take the stress out for them. Just a thought."
— Dad · First user · May 17 2026 · Described the next major feature in one paragraph
100%
Mobile built
Every prompt, every upload, every iteration, on a phone.
1
File to ship
Drag to Cloudflare. Live in under five minutes.
1 URL
Shareable views
Provider, model, and landmark encoded in the link. Copy and share any configuration instantly.
609
Visitors on day one
No paid promotion. No existing audience. Just the tool.
8pm
The ritual
Nightly price entry. By hand. Before the automation arrives.
$20
Monthly cost
The entire infrastructure bill. Claude Pro subscription.
0
Co-founders
Solo. One person. All decisions. All sessions.
The journey
The account opened in July 2025 with a QR code generator and a slot machine game. TokenScale shipped in May 2026. Here is what happened in between.
July 2025
First session ever
QR code generator. Slot machine. Fish identification quiz. All on mobile. The account opens. No plan, just curiosity.
October 2025
The question that becomes TokenScale
A session titled "Pricing model details", why does Haiku cost less than Opus? What are tokens, really? The seed is planted.
October–November 2025
Voice-coding on the move
Conversations happen while driving. Literally. Ideas captured in motion. Building in fragments, between moments, becomes the pattern.
December 2025
Bilton Projects takes shape
The business identity crystallises. Solo freelance setup. The brand that TokenScale will launch under appears for the first time.
April 22 2026
The interactive handbook
A full "Token Economics Handbook" is built, scrolling, editorial, animated. The same educational mission as TokenScale but static. The intellectual ancestor. The moment the idea becomes serious.
Late April 2026
The tool form emerges
Token Auditor → Claude Explorer → Gemini Explorer. The content landmark framing, "what does The Hobbit cost?", appears for the first time. This is the core insight.
May 1 2026
The merge: 16 providers at launch, 22 today
Claude and Gemini versions merge, then expand rapidly across 14 more providers. That 16-provider launch table makes the tool worth sharing. The board later grows again and now tracks 22.
May 13–16 2026
The animation sprint
Slot machine. Candle-master colour reveal. Market banner tinting. Conversation cost visualiser. Price history changelog. Provider-aware theming. The tool becomes the product.
May 17 2026, First real user
Dad's review
"Maybe you should give people clear advice about which AI would be best for them… take the stress out for them.", TokenScale was shared with three people. Dad reviewed it first. In one paragraph, he described the next major feature: the Help Me Choose wizard. TokenScale already had it. He just named why it matters.
May 20 2026, Launch Day
MVP 2.0, Ships · Google I/O fires a starting gun
16,000 lines. 16 providers then, 22 today. 6 nights of manually entered price history. Shareable URLs. The slot machine animation. The learn index. The parchment price history. One HTML file. Live. And on the same day, Google I/O 2026 launches Gemini 3.5 Flash at $1.50/$9 per million tokens. Five times the old Flash price. The tool that tracks AI pricing had a major pricing change happen on launch day. That's the whole point.
May 21 2026
Day two, the retention signal
164 UV. 34% of launch peak. 2.3× the pre-launch baseline. The right number to watch after a launch isn't the spike, it's whether people come back. They did. Price history updated across all 16 launch providers, before the board grew to today's 22. changelog.html ships: a public record of every price movement, in language anyone can read.
May 22 2026
Retrospective ships. The data moat deepens.
This page goes live. Channels that didn't run on launch day, Reddit, Dev.to, email capture, still exist and still will. The data moat grows one night at a time regardless of traffic.
May 23 2026
The UI grows up.
Biggest single-session improvement run since launch. Per-segment sparkline colouring. A sticky topbar with a conversation slider that appears when the hero scrolls away. Quick-nav pill buttons on both the landing page and dashboard. A fully redesigned price history modal, blur backdrop, provider-coloured top stripe, left/right arrow navigation through all 16 launch providers, now 22, plus keyboard shortcuts. Redundant controls removed. GitHub added as a backup destination. The tool is starting to feel like a product.
May 24 2026 · 12:30 AM
Chris's message · The energy lens
A WhatsApp from a sustainability researcher friend. "Hi Will, this looks fascinating. I've been doing a lot of research into the sustainability of AI and the financial stuff is something that I have read about…..", reading between the lines: the energy side of AI had never been connected to the cost side in a way a normal person could see. That gap became the brief. By the next morning: a hidden energy lens, built into the stats bar as an Easter egg. A kettle. A 100W bulb. Both sides of the argument. Built in one day. Found only when you go looking.
May 25 2026
The automation arrives. No pipeline required.
The project notes had a line: "Automation plan, Cloudflare Worker + KV store, nightly cron, manual JSON update initially then scraper." That line is now crossed out. A scheduled AI agent begins by searching the 16 launch-provider pricing pages, a route set that later grows to today's 22. It diffs against yesterday's entries, appends history, takes a dated backup and reports what happened. No Worker. No KV. No server. The discipline that built the data moat, one entry per provider, by hand, every night, becomes infrastructure.
July 1 2026
The drift audit, the watchdog gets watched
A full re-verification of all 22 providers, 66 tiers, against every official pricing page in one sweep, run by Claude Fable 5. Forty tiers exact. Twenty-six with something to answer for: a retired Grok that silently re-bills at 6× its listed price, two model names that never existed, a Llama flagship that was never released, and eight tiers pointing at models their hosts had quietly stopped serving. The uncomfortable lesson: the nightly automation faithfully verified prices, of ghosts. Price-checking isn't existence-checking. Every miss is published in the changelog with dates and final prices, the dead models get headstones in the graveyard, and existence checks join the nightly. A pricing site that hides its own drift has no business reporting anyone else's.
July 2026
From a phone to a system
The phone was the launch story. It is not the operation. What runs TokenScale now is a set of systems built since: a nightly agent that checks 22 provider routes and records each of 66 tier observations as verified or carried; a guarded release pipeline where every build is versioned, integrity-checked, duplicate history dates, dropped files, and card-versus-history drift each fail the build, and kept permanently as a rollback point; a public QA ledger and changelog where every correction is dated and shown, never buried; promo-aware pricing that keeps list prices canonical; a self-tending model graveyard; and anonymous, cookieless analytics. The builds move through Claude Cowork and ChatGPT, working through Codex, one at a time under the same writer's lease and numbered ledger. They are verified and handed over; a human still tests every one and personally ships it. Built by hand on a phone. Kept honest by machines.
July 20–21 2026
The desk drops away
During a hiatus from regular work, the project travels. Sometimes there is a desk. Sometimes there is a phone, a temporary table or a stop on the road. The development environment is no longer a place. The last piece of a rebuilt analytics system is fitted and tested from a phone while travelling.
August 10 2026
The machine grows bumpers
A two-copy verification rule separates the place where work happens from the copy used to judge it. Release leases, immutable version numbers, exact-ZIP testing and preserved rejected builds turn caution into machinery. A correction can no longer quietly rewrite the record it is meant to protect.
August 13 2026
The Pinball Machine gets its name
The question changes from "can AI build this?" to "why am I still routing every ordinary action?" I stop trying to play the ball and start designing the table: triggers, bumpers, gates, boundaries and hard stops. The human still chooses scope and tests every release. The machine learns to move routine work safely between those decisions.
August 25 2026
269 numbered builds later
The Pinball Machine chapter entered the permanent ledger at v709. This story update is v972: 264 consecutive numbered candidates, including the rejected versions the system preserves instead of hiding. The homepage is 48,347 lines, the board is 22 providers and 66 standard tiers, and every ordinary night now has two official-source passes, two watchdogs and a final evidence close.
The point
The nightly price data ritual, one entry per provider, by hand, on a phone, at 8pm, was not a limitation waiting to be automated. It was the discipline that made the data trustworthy. Now the automation has arrived. The habit became the infrastructure.
TokenScale was built to answer a simple question: what does AI actually cost, in words you recognise? The answer, it turns out, is also the story of how it was built.
The launch example · May 2026
The Hobbit cost $0.06 on Gemini Flash-Lite. Claude Opus was $2.85 for the same words. That 47× launch-day spread made the point.
May 20 2026. Product Hunt at 12:01am. Hacker News at 7am PT. X and LinkedIn before noon. Four channels in one day. Here is what actually happened: the good and the honest.
486
Launch day UV
6.9× the pre-launch baseline of 70/day.
164
Day-after UV
34% of peak. 2.3× baseline. People came back.
756
Launch week UV
7-day window. Six times pre-launch equivalent.
95.7%
Cache rate
On peak day. Cloudflare held everything from edge.
$0
Server bill
152 MB served on launch day. Free tier. No invoice.
The honest bit
Product Hunt ended at #82 with 2 points. Hacker News generated real traffic but not a front-page run. No comment thread to engage with. LinkedIn reached animation and entertainment professionals, Will's industry network, not the AI developers the tool was built for. Reddit, Dev.to, email capture, tool directories: none of them happened on launch day. There is no list to email when prices change next time.
"We found that our own hero number, $0.04, was the input cost only. The correct total was $0.06. We corrected it publicly before the traffic arrived. Within 48 hours, Google proved the lesson differently: on Gemini Flash 3.5, the output cost alone is now $1.14."
— The $0.04 → $0.06 correction · May 20 2026 · The better story
Catching and publishing that correction, before the HN post, before the Product Hunt page saw traffic, was the right call. The difference between input cost and output cost is exactly the kind of thing TokenScale exists to surface. Getting it wrong, and saying so, made that point better than getting it right would have.
The launch was not a failure. It was a beginning. Every channel that didn't execute on May 20 still exists. The data history keeps building every night regardless of traffic. The gap between cheapest and most expensive provider is still 35×, and most of the people who need to understand that haven't found the tool yet.
What a colleague saw
Two weeks before launch, Will showed an early version at work. Just Claude pricing at the time, before the tool expanded to 16 providers for launch and then to the 22 tracked today. Before a single marketing post had gone out. When the cost breakdown appeared on screen the room leaned in. Not because they understand tokens. Not because they use the API. They got interested because they could see how the person who built it thinks.
What caught their attention was a document generated after the session: a formal QA training report, formatted for product and engineering teams, based entirely on how Will had worked. Five principles. Seven competencies. A scorecard. Signal-to-noise ratio: 10/10. Acceptance discipline: 10/10. Bug reproduction quality: 9/10. The kind of document you could table in a meeting. It was produced by analysing Will's own prompts and working style across 40 issues raised during the TokenScale build. A person putting their own process under the same scrutiny they applied to the product. That is what landed. The verified dates, the word count badge, the public correction, the before-and-after cost tables: none of these are features to a non-technical audience. They are evidence of an approach. Checking your numbers, being honest when they are wrong, keeping a record. That is a QA instinct. It shows up whether or not you know what a token is.
The signal
"They were not excited about token pricing. They were seeing how the person who built it approaches a problem."
Anyone who has ever had to stand behind a number in a meeting, a budget line, a schedule estimate, a word count, recognises that instinct immediately. TokenScale turns out to speak to that audience too, even if they never open an API.
The message that started it
The energy lens didn't start with a spec. It started with a WhatsApp at 12:30 AM on May 23 2026. TokenScale had been shared quietly with about three people. One was Chris, a friend from university who had spent serious time researching AI sustainability. A voice note. A reconnected friendship. Then his reply:
"Hi Will, this looks fascinating. I've been doing a lot of research into the sustainability of AI and the financial stuff is something that I have read about….."
— Chris · WhatsApp · 12:30 AM · May 23 2026
Reading between the lines: Chris already understood the financial story. What he was living inside was the energy and sustainability side. The two weren't connected anywhere, not on TokenScale, not really anywhere, in a way that a normal person could see at a glance. That gap became the brief.
The kettle
The anchor came from a conversation in 2024: boiling one litre of water in a standard kettle takes about 0.1 kWh. That's enough energy to run somewhere between 30 and 300 standard AI text queries, depending on the model. The kettle versus the query. Concrete, kitchen-scale, something everyone has lived. No jargon required.
0.1 kWh
One kettle boil
One litre of water. The unit everyone already understands.
30–300×
Queries per boil
Depends on the model. Silicon doing matrix math is astonishingly efficient per query.
100W
The bulb
Incandescent. Standardised globally for decades. The watt made physical.
The 100W incandescent bulb was chosen deliberately. Unlike LED bulbs, 4W? 12W? everyone's is different, the 100W incandescent was a globally standardised unit for decades. It is the watt made physical. A trustable comparison.
Both sides
The feature shows two numbers, not one. Same insight as cost: the per-query number looks tiny. The scale number is what changes behaviour.
Query
What this query used
Usually a fraction of a kettle boil. Silicon doing matrix multiplication is extraordinarily efficient compared to heating water molecules.
Scale
A million users a day
That's where the number gets real. The same logic as cost, tiny per query, consequential at scale.
"The feature is hidden by design. It appears only after you've typed a word count, a small ⚡ circle at the edge of the stats bar. It stays quiet until you've done something. Then it's there."
— Energy lens · Built May 24 2026 · One day after Chris's message
Chris is the sustainability researcher who already knows the financial story. The energy lens is the bridge, the thing that connects the two honestly, without alarm, without requiring a PhD. The feature is hidden by design. If you find it, it was meant for you.
Built: May 24 2026. First user: anyone curious enough to type.
Where it started
I got curious recently about exactly when TokenScale started. Not the launch, May 18th, but the actual first moment. The conversation that set it off.
It was April 21st. A single Claude thread about token economics that had ballooned to 100,000 tokens. I asked Claude to explain the compounding cost problem, what happens when every reply has to read the entire conversation history first. It explained it in real time, inside a conversation that was itself compounding. Then I asked it to build an interactive guide about what it had just explained. It did that too.
"The conversation that explained the compounding token problem was itself a compounding token problem. The tool that tracks AI costs was built using AI costs."
— First conversation · April 21 2026 · ~100,000 tokens
April 21st to May 18th. Twenty-seven days from a single runaway conversation to a launched product.
The build is complete. TokenScale will be maintained and updated, prices checked nightly with carried observations marked, new features when they earn their place. But the building phase is done. Whatever it needed to be, it is. · June 1 2026
Phase two · The Pinball Machine
The build was complete. The system was just beginning.
That June sentence was true about the first product. It was not true about what came next. During a hiatus from my regular work in July, I kept building while travelling. Sometimes I had a desk. Sometimes I had a phone, a temporary table and whatever time the road gave me. TokenScale came with me.
The first phase proved that AI could help one person build a real product. The road made the second question unavoidable: could I design the work so the product stayed accurate when I was not sitting there routing every move?
Claude was the original build partner. During phase two, ChatGPT joined through Codex, OpenAI's coding agent. They do not edit the project blindly at the same time. Both work against the same one-writer lease, numbered release ledger and handoff record. That became another Pinball Machine test: could two capable AI systems continue and challenge the same work without splitting the truth?
264
Numbered builds
v709 through v972, including preserved rejections.
48,291
Lines today
About 16,000 at launch. Still one shipping index.
22 / 66
Providers / tiers
The current board, checked against official sources.
Those are not 264 public redesigns. Many are small candidate steps: a test, a correction, a failed idea preserved with its number, or another guard added after the system found a new way to be wrong. That is the point. Phase two is more complicated because it is not only making things. It is deciding what may change, what must be proved and when the ball has to stop.
Official sources22 routes read twice each night
→
Bumpers and gatestests, receipts, watchdogs and hard stops
→
Human judgmentI choose scope, test the candidate and publish
I call it the Pinball Machine because I stopped trying to play every ball and started designing the table. The road did not take me away from TokenScale. It showed me which parts still depended on where I was sitting, and gave me the space to build those dependencies out.
"AI made the developer portable. Systems made the knowledge persistent. The Pinball Machine asks how little routine motion still needs its operator."
The second phase · Built on the road · Still being tested in public
About the creator
Hi, I'm Will Bilton, Bilton Projects, and the person behind TokenScale.
I'm not a traditional software developer. I spent years as a software QA engineer in animation software, then came at TokenScale sideways: I got stuck on a number I couldn't parse. "$5 per million tokens" meant nothing to me until I forced it to mean something: a word count, a kettle, a novel. So I built a tool to do that translation, for myself first.
That's the first half of the story. The more interesting half began once the tool actually worked.
During a hiatus from my regular work, I took TokenScale on the road. The same phone that helped build the launch version became a portable control room. I could work from temporary places, keep the project moving and notice every moment when the system still needed me to route something routine.
The workshop changed too. Claude remained the original build partner; ChatGPT, through Codex, joined the later system phase. The one-writer lease and shared ledger let both agents take turns on the same product without turning the record into two versions of the truth.
Building TokenScale with AI turned out to be the simple half. Keeping it right (prices that move without warning, models that quietly get retired, a pricing page that vanishes behind a login) is the half that never ends. So the question changed. Not "can AI build this," but "why am I still the one doing this?"
I call what came next the Pinball Machine. I stopped trying to be the thing that routes every action and started designing the table instead: checks, triggers, boundaries, recovery paths and hard stops. Since that chapter entered the build ledger, TokenScale has moved through 269 numbered candidates. The product is public. The system around it is the real experiment, and it is still running.
Why keep it public? Because a demo doesn't have to survive contact with reality, and TokenScale does. Prices move, providers rewrite their docs, models disappear, and the record has to hold up anyway. That ongoing pressure is what makes it an honest test of the idea, rather than a nice sentence about it. I'm documenting what I find either way, including the parts where the system still needs me more than I'd like.
Built by Bilton Projects, exploring what happens when the interesting question stops being "how fast can AI write this" and starts being "how little do I actually need to be here."