Claude Opus 5 or Fable 5: Which One Should You Pay For Your Dev Team

Two models from Anthropic, one price gap that decides your budget: Claude Opus 5 costs five dollars per million input tokens, Claude Fable 5 costs ten. A token is the unit of text the model reads and writes, and it's what your bill is calculated on. Fable 5 costs exactly double, for the same quality of context. So the real question isn't which of the two is smarter, but which one is worth paying for your developers, task by task.

Here's the memo I'd send to a CEO who signs off on the bill, not to the dev who writes the code. No technical benchmarks, no charts: cost, turnaround, risk, and a verdict you can apply Monday morning.

The one number that should drive your choice: double

Let's put the prices side by side, because everything starts there. Opus 5 bills five dollars per million input tokens and 25 per million output. Fable 5 climbs to ten in and 50 out. Both models share the same 1-million-token context window (how much code and documentation the model reads in a single pass, the equivalent of a very large repository), and the same 128,000-token output ceiling. The container is identical. Only the price per litre changes.

For equal work, defaulting to Fable 5 means doubling your API bill with no guarantee whatsoever of doubling what gets delivered.

This isn't an accounting footnote. Generative AI's economic potential runs between $2.6 and $4.4 trillion a year according to McKinsey, but that value only lands in your pocket if you keep a grip on the unit cost of each task. A model that costs twice as much and buys you five percent more perceived quality is a bad deal on a dev budget.

What does that come to per developer, per month?

Take a conservative, honest assumption: an AI-augmented developer who burns 50 million input tokens and 5 million output tokens over a month of active work. On Opus 5, the API bill comes out around 375 dollars (250 in, 125 out). On Fable 5, it climbs to 750 dollars. 375 dollars of difference per dev, per month, mechanically, for the same volume of work.

Across a team of five developers, that's nearly 1,900 dollars a month gone up in smoke if you wire in Fable 5 by reflex. That's a junior's salary, every month, against a gain that nothing in day-to-day coding proves.

Look at the last column: Opus 5 is the baseline, the model everything else positions itself around. Fable 5 isn't the standard model at a higher price, it's the frontier model. That distinction is worth money.

What Fable 5's always-on reasoning makes you pay for

That double price isn't arbitrary: it buys reasoning, and that's where the budget trap snaps shut. Reasoning is the time the model spends "thinking" before it answers, producing intermediate tokens that you pay for at the output rate. The more it thinks, the more thorough the answer, the higher the bill and the longer the wait.

On Fable 5, this reasoning is always on and impossible to switch off. You pay for maximum-thinking mode on every single request, including reformatting a file or fixing a typo in some code. On hard tasks a single turn can run for several minutes: great for an architecture problem, ruinous for a quick fix loop where you're waiting to get back to work.

What does an effort level actually mean in practice?

It's the dial that decides how much the model thinks before it delivers. Opus 5 offers five notches: low, medium, high, xhigh, max. You can even turn reasoning off up to the "high" level. In business terms: you set the speed/cost/quality trade-off task by task, instead of having it imposed on you.

That's exactly the lever Fable 5 takes away. Where Opus 5 lets you drop to a low notch for a simple task and climb back up for a hard one, Fable 5 bills full throttle at all times. Worth noting in passing: Opus 5 has a "fast mode" at $10/$50, the same rate as Fable 5, to switch on only when speed genuinely trumps cost.

Opus 5, your team's workhorse

Once you've got that dial in hand, the default choice becomes obvious. Anthropic explicitly positions Opus 5 for complex agentic coding and enterprise work. Agentic coding is when the model plans, writes, tests and fixes on its own across multiple files, instead of spitting out an isolated snippet. That's precisely what an AI-augmented developer does on a real engagement.

The point that changes the maths: Opus 5's low effort levels (low and medium) are unusually capable. In practice, a large share of everyday work runs just fine on a low notch, at a reduced cost and with short latency. That's your main lever for savings, even before the choice of model.

Full disclosure: I bill for AI-augmented senior developers, so I have an interest in these tools living up to the hype. That's also why I watch the API bill closely, not just the demo. In July 2026, across the engagements I run, shifting routine tasks down one effort notch cut the API bill almost in half without the delivered quality budging. The waste wasn't in the model, it was in the setting.

Why start at the lowest effort level?

Because bumping up a notch costs a few seconds, whereas paying the maximum by default costs every month. My rule is simple: I start on Opus 5 at low effort, and I only climb when a task genuinely gets stuck. A senior dev knows how to spot that moment, and that's where at least eight years of experience earns its keep. If you want the full breakdown of API rates and subscriptions, I've laid it out in this comparison of Claude Code pricing, and the question of a team's real velocity is covered in this guide to measuring velocity.

When paying for Fable 5 isn't waste

Everything above assumes normal tasks. When that assumption breaks, Fable 5 earns its place back. It's Anthropic's most capable widely available model, and it targets two specific cases: the most demanding reasoning and the longest agentic tasks, the ones where a run holds up for tens of minutes without going off the rails. A heavy migration, a gnarly architecture overhaul, an algorithmic problem where an Opus 5 miss would cost you more than the extra bill: there, paying double is justified.

Fable 5 also drags along operational constraints a CEO should know about before signing off. Three concrete points.

First, Fable 5 imposes 30-day data retention and isn't available in zero-retention mode. Second, some requests can be refused by its safety classifiers, notably around cybersecurity and biology: a security code audit can end up blocked. Third, those multi-minute turns break the short feedback loop that a fast-shipping team lives on.

Should you worry about the 30-day data retention?

For most projects, no. For a regulated client (healthcare, finance, defence) that contractually requires zero-retention mode, yes: Fable 5 is off the table, and Opus 5 becomes not only cheaper but the only compliant choice. That kind of constraint weighs more heavily on a project's real cost than the price per token, as I detail in this look back at the true budget of AI agents in production.

My verdict: Opus 5 day to day, Fable 5 on the meter

A clear recommendation, the one I apply myself. Opus 5 by default, at low effort, for all the team's agentic coding. You bump up a notch when a task resists, and you only reach for Fable 5 on extreme-reasoning problems or very long runs, treating it as an exceptional expense justified line by line.

How do you apply this verdict starting Monday?

The decision criterion fits in one sentence: if you can't name the specific task that requires Fable 5, then Opus 5 is enough. Paying double "just in case" isn't caution, it's waste you've talked yourself into. Ask your lead dev which model and which effort level your agents run on today: if nobody can answer, you're probably paying full price without knowing it.

"Companies don't want AI. They want to ship faster and spend less. The most expensive model is almost never the answer to that."

Vincent Roye, July 2026

Choosing a model isn't about buying the most prestigious one, it's about industrialising a software production line around the right unit cost. The real edge is in the tuning and the steering, not the premium line in the catalogue. For the purely business angle of this reasoning (staffing, day rates, team structuring), the GoLive Software blog carries the thinking further on the development side.

At Extra Dev, an AI-augmented senior developer (eight years of experience minimum) costs 180 euros a day all in, and I staff them on Opus 5 tuned to the bone, not on the priciest model for comfort. It's that discipline, more than the choice of model, that makes the difference on your end-of-month bill.