Opus 5 Killed the Fable Tax

I've written twice about the Fable Tax — the argument that Claude Fable 5, Anthropic's most expensive model, was never worth its premium for most real work; that you were paying flagship prices for a marginal edge you rarely cashed. I caught some heat for it.

Today Anthropic made the case for me. In its own announcement, it describes Claude Opus 5 as “a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price.” Read that again: the vendor is now selling most of Fable's capability at half of Fable's cost — and at the exact same price as Opus 4.8. The Fable Tax didn't get cheaper. It got refunded.

What actually shipped (the confirmed part)

Price parity is the whole story. Opus 5 is $5/M in, $25/M out — identical to Opus 4.8. (A fast mode runs ~2.5× speed for 2× the price.) It's live today on every platform — API, Claude Pro and Max, Cursor, GitHub Copilot — it's the new default on Claude Max and the strongest option on Pro, and it succeeds Opus 4.8 from May.

The dial is the real feature. Opus 5 ships an effort setting — low, medium, high, plus a max mode — that lets you optimize for intelligence or conserve tokens per turn. Fortune put it on the headline: a toggle between cost and capability. That's the tradeoff as a knob, not a model you have to pick.

Alignment up, offense down — on purpose. Anthropic calls it its most aligned model to date (its lowest misaligned-behavior score yet) with stronger cyber-defense guardrails than 4.8 — and it's deliberately weaker at offensive cyber, sitting behind on exploit-development benchmarks by design. When it refuses on safety grounds, the API now automatically falls back to Opus 4.8 instead of throwing an error.

Less hand-holding. It needs fewer round-trips, self-verifies its work, and recovers from its own errors. The API also added automatic fallbacks and mid-conversation tool swaps that don't invalidate the prompt cache (both beta). One ops note on that auto-fallback: it means the serving model can change mid-workflow — so log it and surface it in your traces. Invisible model-switching is how a clean demo becomes a haunted production system.

What's still rumor (the honest part)

This is where the coverage runs ahead of the facts. Anthropic was unusually quiet — no splashy model card, no pricing-page fireworks — so treat these as reported-but-not-confirmed: a 1M-token context window (widely repeated, matches the 4.8-era window, but not spelled out in the announcement), the exact name of the max effort tier (community reports say xhigh), and any “beats GPT/Gemini” screenshot going around. If a claim traces to one X post, it's a rumor, not a benchmark. The model reportedly first surfaced as a stealth build codenamed Honeycomb, flickering into Cursor's picker around July 9 before being pulled within hours — a classic Anthropic soft-launch.

The dial is the actual news

Everyone will fixate on “near Fable at half price,” and fair enough. But the thing that changes how you work is the effort dial. The old mental model was a ladder of models: Haiku for cheap, Sonnet for balance, Opus for hard, Fable for the frontier — a different price per rung, a decision every time. Opus 5 collapses most of that ladder into one model with a throttle. For most teams, “which model?” becomes “how hard should it think about this one?” — capability becomes routable per task instead of rented at full intensity by default.

A working policy:

The rule is simple: start lower when a mistake is cheap and obvious, start higher when it's expensive or subtle. If a low-effort answer fails inspection, bump that turn — don't re-run the whole conversation at max because one step needed it. (And ignore anyone selling a clean “effort tier = cost multiplier” formula; Anthropic didn't publish one, so they're decorating guesses with decimal points.)

Hot take: the dial will save a serious team more than the token price ever could, because it kills the expensive habit nobody admits to — spending maximum intelligence on work a deterministic script could almost do.

The benchmarks (Anthropic's own — trust, but verify)

Every number below is Anthropic-reported, from the announcement; independent results will take a week or two. But they're specific, not vibes:

Note the shape of the coding claim, too: a model can cost the same per token and still be cheaper per finished task if it needs fewer retries and less babysitting. Opus 5 starts at 4.8's exact price and, on these numbers, gets there with less — which is why this launch is ugly for 4.8, not just for Fable.

Discount vendor benchmarks 20% if you like. “Within 0.5% of Fable at half the cost” and “beats Fable on computer-use at a third the cost” are the two lines that gut the Fable argument — and they're Anthropic's own lines, not mine.

Opus 5 vs Opus 4.8 vs Fable 5

Who should still touch Fable? Almost no one.

Here's the honest part, because a one-stop-shop that only dunks isn't useful. Anthropic still recommends Fable 5 for the most advanced, long-horizon autonomous agentic work — the stuff where you hand a model a big goal, let it run for hours with minimal supervision, and the last few percent of reliability compounds over thousands of steps. If that is literally your production workload, Fable still edges it.

For everyone else — every chat, every code review, every agent a human still checks, every normal build — Opus 5 at half the cost is the default worth testing first, and paying the Fable premium becomes a choice you have to justify with your own numbers, not a default you fall into. Hot take: the premium tier's job used to be “best available.” Now its job is “the sliver you'll almost never need.” That's the Fable Tax, gutted.

What to do Monday

1. Switch your default to Opus 5. Model id claude-opus-5. Same price as 4.8, better on Anthropic's numbers, near-zero migration cost. Little reason to wait — and 4.8 mostly lingers now as the safety-fallback target, not a place to stay.

2. Learn the dial. Run routine work at low or medium effort — cheap and fast — and crank it to high or max only when the problem earns it. That's where the real savings live: most of what you do doesn't need maximum reasoning, and now you don't pay for it.

3. Measure finished tasks, not benchmark theater. Anthropic's numbers are a strong prior; your workload is the actual test. Run Opus 5 against your current 4.8 and Fable flows on the same inputs, and track success rate, retries, human corrections, and total cost per accepted result — not tokens on a leaderboard.

4. Stop paying the Fable tax. Unless unsupervised long-horizon agents are your core workload, drop Fable to the break-glass tier and default to Opus 5. For the record, I switched every machine and project I run to claude-opus-5 as the default today — global, work, and every side project.

The pattern here is the one that's held all year: the frontier doesn't stay expensive, it becomes the new baseline. Fable was the frontier for a few months. Now near-frontier capability costs what the workhorse cost yesterday, with a dial bolted on so you only pay for the thinking you actually use. The move was never loyalty to a tier — it's switching the moment the math changes. Today the math changed. Opus 5 is the default now. Don't be the last one still paying the Fable tax.