Anthropic’s Best AI Model Just Became a Metered Utility

Anthropic’s flagship moved to a pay-as-you-go meter for standard plans — a preview of where the whole market is heading. The question is no longer which model is smartest. It’s which model is right for the job, at a price that makes sense.

HEADLINE STORY

Anthropic pushed back the deadline twice before finally settling how you’d pay for its most powerful model. The answer landed last week, and it’s a preview of where this whole market is heading.

Here’s what happened, in plain terms. Anthropic’s top model is called Fable 5. It’s the smartest thing they make, and it’s costly to run. For a while it was bundled into paid Claude subscriptions at no extra charge. Starting July 20, that stops for the standard plans. The top-tier plans keep it included, at up to half of a plan’s usage: Max, at $100 to $200 a month, and Team Premium. This time Anthropic set no end date, which the earlier windows all had.

If you’re on a standard paid plan, Fable 5 moves to a pay-as-you-go meter, billed by the token. A token is a small chunk of text, and the model is charged for every one it reads in and writes back: $10 per million tokens in, $50 per million out. To soften the landing, those users get a one-time $100 credit to burn through first.

Translate the meter into something familiar. A million tokens is about 750,000 words — a lot of drafting, summarizing, and back-and-forth. But heavy users can consume that capacity faster than they might expect. Long documents, deep research, coding, large context windows, and repeated analysis all increase the bill. That rate, by the way, is double what Anthropic charges for its next model down. The flagship is no longer the model you reach for without thinking. You use it when the work justifies the added cost.

Anthropic says this is about compute capacity, not strategy: demand for the model outran the hardware they had to serve it, and that’s probably true. My read is that the final call was also competitive. Rivals have gotten good enough that pulling the best model out of subscriptions entirely would have sent people shopping.

What Else Mattered

A rival drew even, at a lower price

OpenAI’s newest flagship, GPT-5.6 Sol, now scores essentially the same as Fable 5 on an independent capability index, at half the price on the text it reads in and less on the text it sends back. By one independent tally it runs about a third the cost per finished task. When the second-best option is close enough in quality and materially cheaper, “just use the best model” stops being obvious advice.

The launch that keeps slipping

Google’s Gemini 3.5 Pro was widely reported to be arriving July 17. That date came and went with no public release, no model card, and no price, and it still isn’t out. Worth being precise: the only timing Google ever put on the record was a vague “next month” back in May, and every date since has been reported or rumored, not promised. The takeaway is the useful part. Don’t hold a current decision hostage to a product that has not been announced, let alone shipped. Make today’s decision using the products that actually exist today.

The floor keeps dropping

Between newer low-cost models from xAI and open-weight releases landing almost weekly, the price of “good enough for most work” keeps falling. The gap between the premium tier and the commodity tier is widening on price and narrowing on quality at the same time.

The Pattern Behind the Week

For two years the AI question was simple: which model is smartest? That question is closing. The models at the top are now close enough to one another that, for most business work, the difference doesn’t show up in the result. It shows up on the invoice.

It’s no longer “which model is best.” It’s “which model is right for this specific job, at a price that makes sense.”

So the real question has quietly changed. It’s no longer “which model is best.” It’s “which model is right for this specific job, at a price that makes sense.” The market is sorting itself into tiers: a premium tier you pay up for, a middle tier that handles the bulk of the work, and a cheap tier for high-volume routine tasks. The operators who come out ahead will be the ones who match the job to the tier instead of defaulting to the most powerful, and now most expensive, option for everything.

That’s not a technology skill. It’s a purchasing discipline, and you already run it everywhere else in your business. You don’t put premium fuel in every truck, and you don’t special-order top-grade material for a job a standard grade handles fine. Same logic, new line item.

What To Do Now

1. Find where you’re paying top-tier prices for routine work. Most of what a mid-market team does with AI — first drafts, summaries, cleaning up notes, quick research — doesn’t need the smartest model. Route that work to a mid-tier model and reserve the flagship for unusually complex work or situations where a weak answer carries real consequences: a first-pass contract analysis, a thorny operating question, a decision memo that deserves your strongest analytical tool.
2. Know which plan your team is on before the week is out. If your people lean on Claude’s Fable 5 daily and you’re on a standard plan, that $100 credit is a runway, not a subsidy; once it’s gone, you’re paying metered rates. Decide now whether the work justifies moving up to Max or changing how you route it.
3. Write down a one-line rule. Something like: “Use a top-tier model only when the work is unusually complex, or a weak answer could cost us real money or real trust.” That one sentence is your cost-discipline rule, and it will save you more as your usage grows than any tool you could buy.

What I’m Watching

Whether Gemini 3.5 Pro actually ships this time, and what it costs when it does. And whether AI model selection becomes a normal part of departmental budgeting, the way freight and software licensing already are.

One Question Worth Asking

If you had to name the one task in your business that needs your most expensive AI — and the ten that don’t — could you do it right now? The operators who can answer that are likely to spend less — and get more value from what they do spend.