The cheapest token is the one you never spend

I spent about six weeks last quarter watching a bill climb for reasons nobody on the team could explain.

The setup was ordinary. We had a research agent that read a pile of vendor documents, pulled out claims, checked them against each other, and wrote a recommendation. It worked. People liked it. Then someone in finance asked why a tool that produced maybe forty memos a month cost more than the humans who used to write them.

The answer was embarrassing when we finally traced it. We had wired the whole pipeline to a single frontier model because that was the one that worked when we built the prototype. Every step used it. Reading a PDF to see whether it was a contract or an invoice used it. Renaming a file used it. The step where the agent actually had to weigh two contradictory claims and decide which one held up also used it, and that step was maybe four percent of total calls.

We were paying frontier prices for filing.

Reasoning is not a setting you turn on for the whole app

The mental model that got us into trouble is a common one. You pick "the best model," you point your agent at it, and you assume that better answers per call add up to a better product.

They don't add up that way, because most calls in an agent workflow are not interesting. A long-running agent spends most of its life doing clerical work. It fetches, parses, classifies, reformats, retries, summarizes, and hands off. Those steps have a right answer that a cheap model finds just as reliably as an expensive one. Paying more does not buy you a more correct filename.

What paying more buys you is judgment on the steps where judgment is scarce. There are usually few of them, and they usually arrive late, after the agent has gathered enough context that the next decision actually costs something to get wrong.

So the useful question is not "which model is best." It's "where in this workflow does a shallow answer become expensive?" Answer that and the routing designs itself.

Anthropic's own framing points the same way

The Fable 5 launch is a decent illustration, mostly because of what Anthropic said about where the model's advantage shows up.

Anthropic describes Fable 5 as a Mythos-class model made safe for general use, and says its capabilities exceed any model the company has made generally available. The part worth reading twice is the qualifier attached to that claim. Anthropic says the model's lead grows as tasks get longer and more complex. Not that it wins everywhere. That it wins more as difficulty rises.

Read that as a routing instruction rather than a marketing line and it tells you something concrete: on a short, easy call, the gap between the expensive model and the cheap one is small, and you are paying for a margin you cannot observe. On a long, tangled one, the gap is where your product quality lives.

The reported results line up with the framing. Anthropic says Stripe used Fable 5 to complete a codebase-wide migration in a 50-million-line Ruby codebase in a day, against more than two months of manual work by a team. Anthropic says the model scored highest among frontier models on Cognition's FrontierCode evaluation even at medium effort, and posted the top score on Hebbia's Finance Benchmark for senior-level reasoning. Anthropic also says it stays focused across millions of tokens on long-running tasks and improves its output using its own notes.

Every one of those is a hard, long, expensive-to-get-wrong task. None of them is "classify this PDF."

The price tag is the design constraint, not a footnote

Anthropic lists Fable 5 at $10 per million input tokens and $50 per million output tokens.

Agent workflows have a nasty property that makes those numbers land harder than they look. Token usage doesn't scale with the number of tasks you run. It scales with the number of turns, and turns multiply. An agent that reads twelve documents, cross-checks them, revises once, and validates its own output has spent far more than twelve calls. Loops, retries, and self-checks are where the real spend hides, and those are exactly the mechanics that make agents useful in the first place.

So you get a squeeze. The features that make an agent good at long work are the same features that make an expensive model expensive to run inside it.

There are two honest ways out. Use the expensive model less, or use it only where it changes the outcome. The second one is the same as the first, but with a rule attached.

What a layered workflow looks like in practice

After the finance conversation we rebuilt the pipeline around a simple question asked at each step: if this step returns a mediocre answer, what does it cost us?

Steps where a mediocre answer costs nothing got the cheapest model that passed. Classification, extraction, routine summaries, the retry path. Steps where a mediocre answer costs a little got a balanced model. Drafting, structuring evidence, the ordinary middle of the workflow.

Steps where a mediocre answer costs real money got the expensive one. In our case that was exactly two places. The step where the agent has to reconcile sources that disagree, and the final pass before a human puts their name on the output.

Our bill dropped by a bit more than half. Quality on the memos went up slightly, which surprised me until I thought about it. When you route deliberately you are forced to write down what each step is for, and half the wins came from that rather than from the models.

The platform we run this on, Buda, treats that layering as the normal way to build rather than an optimization you bolt on later. Fable 5 sits there as a subscription-only premium tier rather than the default, which is a design opinion I've come around to. The credit multipliers make the tradeoff legible instead of abstract: Sonnet 4.6 runs at 1.0x, Opus 4.8 at 1.7x, Fable 5 at 3.3x. When the cost of a choice is visible at the moment you make it, you make it differently. That is most of the value. The rest is that a persistent workspace means the cheap steps and the expensive step are looking at the same context, so routing doesn't cost you continuity.

The fallback nobody plans for

One detail worth knowing if you are building on Fable 5, because it will confuse someone on your team at 2am.

Anthropic says the model includes safety classifiers, and that some cybersecurity, biology, chemistry, and distillation-related requests may automatically get a response from Claude Opus 4.8 instead. Anthropic puts this at under 5% of sessions on average, with more than 95% involving no fallback.

Under 5% is small until you are running thousands of sessions and someone opens a ticket about inconsistent output on a security review workflow. It is not a bug and there is no flag to flip. The practical response is to define the business task narrowly enough that the classifier isn't ambiguous, and to keep a human in the review loop on the categories where it might fire.

Where I've landed

The strongest model available is a bad default and a good tool.

I don't think this is a permanent state of affairs. Prices fall, cheap models get better, and some of this routing discipline will look quaint in three years. But right now the gap between tiers is wide enough that spending it evenly across a workflow is just waste, and narrow enough on easy tasks that nobody notices the waste until finance does.

If you run agents and you have never looked at which step is burning your budget, go look. My guess is it's the filing.