What a Token Turned Out to Be Worth

TL;DR

I promised a boring null result if that’s what I found. Here it is. Time of day, parallel sessions, and how full my usage bar was made no difference I could find to cost or speed.
The model is the only lever. The same completed task cost $0.19 on Sonnet, $0.21 on Opus, and $0.58 on Fable. Everything else I measured was noise.
There is no gradual squeeze, just a cliff. Performance was identical at 99% of my weekly limit as at 5%. Then a hard 429, with no warning.
My subscription buys tokens at about three cents on the dollar. A full weekly bar is worth roughly $1,400 in API terms. Run bunx ccusage and you can see your own number.
People do this properly, at scale. OpenRouter, Artificial Analysis, and Anthropic themselves publish far better versions of what I attempted. Links below.

In July I wrote that I couldn’t tell you what a token was worth on my own subscription, and that I was going to measure it instead of wondering. I promised that if the answer was a boring null result, I’d say so plainly.

The answer is a boring null result. Mostly.

What I Actually Ran.

The setup was simple on purpose. One small, fixed coding task against a frozen slice of a codebase, run over and over by claude -p on my own Max plan. Same prompts every time, scored by actually running the tests the model wrote against. 419 runs across three models (Sonnet, Opus, and Fable), covering all 24 hours of the day, with at least a dozen runs in every hour. Alongside each run I logged the usage bars, how many real sessions I had going, and every throttle signal. About $146 of API-equivalent spend went into the probes themselves.

That’s a fraction of what a real study would run, on a sample size of one account. Keep that in mind for everything below.

Night Looks the Same as Day.

My hunch in July was that the same task might draw less from my window at 2am than at 2pm. It doesn’t. Normalized cost across all 24 hours sat between 0.91 and 1.12 of the median, with no pattern. If anything my night runs were a whisker more expensive, which is the opposite of the off-peak discount story, and well inside the noise for this sample size.

This makes sense once you think about where cost variation on a fixed task even comes from. The per-token price never moves. What moves is how many tokens the model happens to spend, and that turned out not to care what the clock said.

Ten Parallel Sessions Cost the Same as One.

The other hunch was that the first thread in a session was the cheapest one, with a premium on every thread after it. Also no. I compared probes that ran while I had zero to two real sessions going against probes that ran alongside ten or more. Cost sat at 0.99 of the median in the quiet bucket and 1.04 in the busy one. Throughput and time-to-first-token didn’t move at all.

I never induced the load, though. That comparison rests on the natural rhythm of my own workday, not on a controlled burst of parallel probes. The fixture for that test exists and never ran. So read this as “no effect I could find,” not “proven flat.”

There Is No Squeeze. There Is a Cliff.

This one surprised me. I expected that as my weekly bar filled up, something would degrade. Slower tokens, longer waits, subtle rationing. Nothing. At 76 to 99% of the weekly limit, cost and speed matched an empty bar.

Then, one August morning, the bar hit 100% and back-to-back runs got a hard 429: “You’ve hit your weekly limit.” No warning, no taper, no grace, at least none my instrumentation could see (a 429 the CLI quietly retries through leaves no trace in my logs). The line doesn’t bend. It just ends. If you plan work around your limits, that’s worth knowing. You will get full service right up to the moment you get none.

The Only Lever Is the Model.

The same completed task, passing the same tests, cost triple depending on which model answered.

Sonnet cost $0.187 at 61 tokens per second. Opus cost $0.210 at 51 tokens per second, and took fewer turns to finish, so its higher sticker price netted out to a 12% premium. Fable cost $0.582, three times Sonnet, and was also the slowest to start responding.

Every knob I suspected the platform was turning behind my back turned out to be still. The only dial that matters is the one I control myself, on every single request. I spent weeks looking for a hidden variable and found the visible one.

What the Subscription Is Actually Worth.

The bars finally have an exchange rate. Across more than a hundred measured reset windows, a full five-hour session bar works out to about $206 of API-equivalent usage. A full weekly bar is about $1,400. Treat those as estimates: I switched between accounts during the window (I have three), and the scripts had to throw out the transition periods. A Max subscription week costs about $46. That’s roughly three cents on the dollar against API pricing, if you actually use the thing.

And I use the thing. You can see this yourself even on a flat subscription: run bunx ccusage and it reads your local Claude Code transcripts and prices them at API rates. Mine says $22,000 since June 2, across those accounts. I pay flat monthly fees and consumed twenty-two thousand dollars of raw material at sticker price. That gap is the entire answer to why subscriptions have bars, and it’s also why I stopped resenting the bars.

People Do This Properly.

This was a small, scrappy version of research that other people do at scale, with real sample sizes, across providers. If the question interests you, their work is better than mine.

OpenRouter sits in front of hundreds of models and publishes live rankings, token market share, and per-provider throughput and latency. Artificial Analysis runs continuous independent benchmarks on cost, speed, and quality across every major model. And Anthropic’s own Economic Index publishes aggregate data on how these tokens are actually being used across the economy. My 419 runs are a rounding error next to any of them. What they can’t do is run on your account, on your plan, on your work. That part you have to measure yourself, and it costs an afternoon of setup.

Closing the Ledger.

I opened this series worried the line was quietly moving. It isn’t, or at least it didn’t move on my account, in my window, on my task. I was wrong about the hidden dials, and I’m glad I checked instead of letting the suspicion calcify into a belief.

What I got instead was more useful than a conspiracy. I know what my subscription is worth in dollars now. I know the limits fail loudly instead of degrading quietly. And I know the only pricing decision that matters is one I make consciously, every time I pick a model. The carpenter finally knows what the lumber costs. It turns out the lumber yard was honest, and I just needed to read my own receipts.

David Kerr is the founder of Kerrberry Systems. He builds custom software for businesses that want a straight answer about what their tools actually cost, not just what the sticker says. Find him on LinkedIn or GitHub.