Claude Opus 5 vs Fable 5: We Couldn't Build a Task That Separates Them

Anthropic sells two models above the Sonnet tier:

Anthropic's own framing when Opus 5 launched was that it "comes close to matching Fable 5's capabilities at half the price." That's an unusually modest claim to make about your own flagship, and it invites an obvious question: close enough for what?

We tried to answer it empirically. It didn't go the way we planned, and the honest write-up of that is more useful than a manufactured verdict.

The easy battery was useless, so we made it harder

Our standard tasks — extraction, a five-constraint scheduling puzzle, a two-bug debug — are already passed by every GPT-5.6 tier including the budget one. Running them here would have produced a tie that meant nothing.

So we wrote three harder tasks, each still gradable by script:

1. A bug that hides on obvious inputs. An interval-merging function that overwrites the previous end instead of taking the max — correct on typical input, silently wrong when one interval is fully contained in another. A pass requires spotting it, producing a triggering input, and fixing it with max().

2. Letter counting. How many times does "r" appear in "strawberry raspberry refrigerator"? The answer is 10. Language models are famously bad at this because they see tokens, not characters.

3. Two simultaneous hard constraints. Exactly four sentences about rain, every sentence exactly five words, and the word "rain" never used. All three conditions machine-checkable.

Both models passed all three

The letter-counting result surprised us most. That task has been a reliable way to embarrass frontier models for two years. Both got 10.

The dual-constraint task is the one we expected to break at least one of them, because it requires holding two independent rules while generating. Both held.

What the premium actually buys: fewer words

Two things worth noting.

Fable 5 is more concise — 26% fewer output tokens for the same verdicts. Because you pay per token, that erodes its own price premium: the list ratio is a clean 2× on both input and output, but the measured cost ratio is 1.49×.

Fable 5 was also faster, by about 11%. Same cause as the Luna/Terra inversion we found in the GPT-5.6 tiers, running the opposite direction: fewer tokens, less time. The premium model is not the slow one here.

So the premium is real but smaller than advertised, and it buys terseness rather than correctness — at least on anything we could construct.

The constraint test, and an honest note about variance

We ran the no-letter-"e" test five times per model:

Fable ahead — except we also ran this exact test against Opus 5 two days earlier, where it scored 1/5. Across both runs Opus 5 is 4 of 10.

That spread, 1/5 and 3/5 on identical prompts, is the real result. It isn't a ranking; it's a warning that a five-trial sample on this task can't distinguish a 40% model from an 80% one. Anyone reporting a single five-run comparison as a verdict — including us, if we'd stopped after one — is reporting noise.

What survives: neither model reliably honors a hard output constraint, and the failures are near-misses. Opus 5's worst run used five instances of the forbidden letter; most failures used exactly one.

What we can and can't conclude

We could not find a task where the premium pays for itself. On three deliberately hard, objectively gradable problems plus a constraint test, Opus 5 matched Fable 5 everywhere except a lead inside our own noise floor.

That is not proof they're equivalent. It's the limit of our battery, and the limit is specific: every task here resolves in one turn and a few seconds. Fable 5's positioning is long-horizon agentic work — tasks that run for many minutes across many tool calls, where small per-step advantages compound. Nothing in a single-turn battery can measure that, and we won't pretend otherwise.

The practical read: if your workload is request-shaped — one prompt, one answer, seconds not minutes — the evidence here says Opus 5 does the same job for about two-thirds the real cost. If your workload is a long autonomous agent, our data has nothing to say about it, and Anthropic's own materials point at Fable 5.

And validate constraints regardless of tier. The most expensive model available failed a trivially checkable rule once in five runs.

Run it yourself

No account or API key needed:

• The letter-counting task — answer is 10
• The dual-constraint task — four sentences, five words each, no "rain"

To run both models side by side and see the token counts behind the cost, that's Promptster.

Run 2026-08-04 against claude-opus-5 and claude-fable-5 via the Promptster API, max_tokens: 4000, default temperature (both models reject an explicit temperature). Costs computed from list pricing: Opus 5 $5/$25, Fable 5 $10/$50 per million input/output tokens. The three graded tasks were run once each — sufficient for pass/fail against a deterministic answer, not sufficient to characterise latency or token variance. The constraint task was run five times per model, and as discussed above, five trials is not enough to rank two models on it.