The Free Window on Anthropic's Newest Model Just Closed. Now That It's Paid API-Only, Here's the Bill Math.
Two weeks ago I wrote that Anthropic had shipped its best model to the public for free, with a meter coming, and that the smart move was to spend the window learning what the model could do rather than wiring it into anything load-bearing. The window is now closed. Fable 5 is a paid API tier ($10 per million input tokens, $50 per million output), and the consumer free access on the Pro and Max plans has ended.
If you took my advice, you spent the free run experimenting and you owe nothing. If you did what's more natural (built the free model into a workflow because it was good and it was there), this is the week the invoice starts.
What changed since June 11
A quick recap for anyone who didn't follow the saga, because it had a wrinkle.
Fable 5 launched free on June 9. The original plan was a short free window (call it the trial) after which usage would be metered at standard API rates. Then the model got pulled offline by a government-forced shutdown that ran from June 12 to June 18, which I also wrote about at the time. That outage effectively stretched the free period: the days the model was unavailable weren't days anyone could burn through a trial. So the practical free run lasted longer than the original "about two weeks" implied.
That extension is the only genuinely new fact here, and it cuts against you in a subtle way. A longer free window means more people had more time to build more on top of a model that was free: which means more workflows quietly flipped from "costs nothing" to "costs $10 in, $50 out per million tokens" this week, often without anyone re-checking the math.
The number that matters is output
Look at that price pair again: $10 input, $50 output. Output costs five times input. That ratio is the whole story for your bill, and it's the part people consistently misjudge.
It means the cost of a Fable 5 workflow is dominated by how much the model writes, not how much you feed it. A task that reads a large document and returns a one-line answer is cheap. A task that reads a little and generates a lot (drafting, code generation, long structured responses, anything verbose) is where the money goes. If you've got an agent that thinks out loud across many turns, every one of those tokens is billed at the expensive rate.
So before you let anything run metered, the question isn't "how big is my prompt." It's "how much is this thing generating, on every call, times how often it runs."
Where the spend hides for a solo operator
The dangerous costs aren't the ones you watch. They're the ones on a loop.
A chat feature you use by hand has a natural ceiling: you can only type so much. The things that generate a surprising bill are the automated ones: a background agent that fires on a schedule, a batch job that processes a queue, a "summarize every new item" hook that you set up during the free window and forgot was running. Those don't wait for you to be paying attention. They ran for free for two weeks and they'll keep running, now billed, until you look.
This is also the week to remember that Anthropic moved Agent SDK and headless usage onto a separate metered credit pool earlier this month: a fixed monthly allotment on the Pro and Max tiers, then API rates beyond it. If your automations lived under that umbrella, "free model" and "free to run the automation" were two different clocks, and both have now stopped.
The honest counter-take
Paid isn't the same as expensive, and I don't want to scare you off a model that's genuinely good. For interactive, human-in-the-loop work (you asking, the model answering, you reading), Fable 5 at these rates is cheap relative to your hourly value. A few dollars a day to have the best model in the loop while you work is not the line item that decides your month.
The risk was never the price per token. It's unattended volume: the looping job, the over-eager agent, the workflow that generates ten times more output than the task needs. Metered pricing is only painful when something runs a lot without you watching. If everything you do with Fable 5 has a human at the keyboard, you can mostly stop worrying and just use it.
What I'd actually do this week
Three moves, in order. First, list everything you wired to Fable 5 during the free run and flag anything that runs on a schedule or a trigger rather than a keystroke: those are your cost centers, not your chat window. Second, for each of those, look at output volume specifically, because output is the 5x line; cap max tokens, trim verbose system prompts, and kill any "think out loud" behavior that isn't earning its keep. Third, set a billing alert at a number that would annoy you, today, so the first signal of a runaway job is an email and not a statement.
Then use the model. It's good, and at human-in-the-loop volumes it's worth it. Just make sure the things spending your money are the things you'd choose to spend it on, not a hook you set up two weeks ago when none of this cost anything.
Author
Lukas
@lukcombinator