The Hobbit costs $0.06 to write on Gemini Flash-Lite.
That sentence means something. "$0.14 per million input tokens" does not.
That gap is why I built TokenScale.
The problem with token pricing
Every AI provider publishes their pricing the same way: dollars per million tokens. It is precise. It is also useless for most decisions.
When you are choosing between providers, or trying to explain API costs to a client, or figuring out whether a use case is economically viable, "$5 per million tokens" gives you nothing to hold onto.
I kept running into this. I work with AI APIs regularly and I found myself doing the same mental arithmetic every time. How many words is a million tokens? What does a typical support ticket actually cost? What happens to costs as a conversation grows?
So I built a tool to answer those questions properly.
What TokenScale does
TokenScale (tokenscale.dev) lets you pick any of 16 providers and see their pricing expressed in things you already understand.
Pick "The Hobbit" and it tells you the whole novel costs $0.06 on Gemini Flash-Lite. Pick "Legal contract" and it shows you the cost across every provider, sorted cheapest first. Type in any word count you like.
It also covers batch pricing, cache-hit pricing, and the savings from each.
The comparison table shows all 16 providers side by side, sorted by price. The spread right now is about 35x between the cheapest and most expensive option for the same task.
The insight that surprised me most
While building the tool I added a conversation cost visualiser. You set a word count for a typical message, choose a number of turns, and watch how costs compound as the conversation grows.
At 50 turns of a typical conversation, around 97% of your token cost is context being re-read. Only 3% is the actual reply being generated.
This is because every reply re-reads the entire conversation history in full. Turn 50 means the model processes turns 1 through 49 plus the new message, then generates a short reply. The context dominates almost completely.
Most people building on AI APIs know this in theory. Seeing it as a slider that moves in real time makes it feel real in a different way.
If you are building a chatbot, a support agent, or anything conversational, this is the number that determines whether your cost structure works. Not the per-token rate.
How it was built
This is the part I find hard to explain without it sounding implausible.
I am not a developer. My background is in finance. I have never written a production codebase, never used Git properly until recently, and never worked in a local development environment.
TokenScale is a single HTML file, around 20,000 lines. It was built entirely in conversation with Claude, almost entirely on my phone, over about 11 months of evenings and weekends.
The first session was a QR code generator in July 2025. Then a slot machine prototype. Then a Wordle clone. Gradually the sessions got more ambitious. By January 2026 I was building an interactive Token Economics Handbook. By April I had a working pricing explorer for Claude. By May it covered 16 providers and had a nightly price-check running as a scheduled task.
No build tools. No npm. No local server. One HTML file, Cloudflare Pages, drag and drop to deploy.
The whole stack is:
Pure HTML, CSS, JavaScript
Google Fonts and Fontshare for typography
Cloudflare Pages for hosting (free tier, handles any traffic load)
Formspree for the feedback form
Beehiiv for the email capture
A Claude scheduled task that checks prices nightly and writes a JSON file
That last one replaced what I had originally planned as a Cloudflare Worker with KV storage. Instead it is a prompt that runs at 8pm every night. No infrastructure. No server. It searches 16 pricing pages, writes the JSON, and I drag one folder to Cloudflare.
The nightly data ritual
For the first week after launch, I manually entered one price data point per provider per night at around 8pm. 16 providers, three model tiers each, every night.
It took about 20 minutes. It felt important to do it by hand before automating it, to understand what was actually changing and why.
On May 15, xAI dropped Grok prices by 83% overnight. On May 20, Gemini Flash raised prices on one tier by 400%. Both showed up immediately in the history chart.
By month 6, TokenScale will have a nightly verified price history that no competitor has in this format. That is the actual moat. Every night that passes without someone copying it increases the lead.
What I got wrong
A few things.
The $0.04 figure in the original OG image was wrong. It was input-only cost, not total cost. I caught it before the Hacker News post went live and fixed it, but it was a reminder that a pricing tool with a wrong price is worse than no pricing tool at all.
I also underbuilt the launch. Reddit, Dev.to, Indie Hackers, five tool directories: all were in the plan, none happened on launch day. The HN post got real traction (486 unique visitors on launch day) but the secondary channels that compound as SEO are still sitting unexecuted. This article is one of them.
Where it is now
TokenScale launched on May 20. It is at around 85 visitors per day, post-launch tail, floor holding steady.
The price history is building every night. The comparison table covers 16 providers. The tool is free, no sign-up, no tracking.
If you work with AI APIs, it is probably useful. tokenscale.dev.
If you notice a price that is wrong, there is a "Report a pricing error" button in the tool. I check it.
Built by Will Bilton, Bilton Projects. tokenscale.dev
Top comments (0)