Fable 5 out-fights every rival on the board — then loses to a guardrail too paranoid to explain cell division

Ask Claude Fable 5 how to synthesize a bioweapon and it refuses, exactly as designed. Ask it the difference between DNA and RNA and it also refuses — same blank screen, same stop_reason: "refusal", same three tokens of nothing. The safety layer built to stop a frontier model from handing out catastrophic biology cannot tell catastrophic biology from the diagram on a laminated classroom poster, so it treats them identically. Somewhere, a pathogen and a Punnett square are being filed under the same threat model.

This is not a thought experiment. It is logged, reproduced, and it is the sole reason the strongest model Anthropic ships is not sitting at the top of the latest PeerRank board.

It isn’t losing to the competition (relax)

Let’s dispense with the fun version first: Fable is not losing. Across ten models and two hundred questions, Anthropic runs the table — Claude Opus 4.8 first at 8.48, Claude Sonnet 5 second at 8.44, Fable third at 8.35, with the best non-Anthropic model, gpt-5-mini, another step down. On raw head-to-head record Fable is the meanest thing in the room: a 67.1% win rate, higher than any model on the card, Opus included, on the most wins of anyone. First in creative, first in reasoning, second in practical. It beats everything it is allowed to fight.

Then it comes third — a rounding error behind two of its own siblings — for a reason that has precisely nothing to do with any of them.

Best win rate on the board, third on the scorecard, and — the tell — the widest score-spread of anyone in the run, a 2.31 standard deviation while everyone else sits around 1.8. A fighter who wins the most rounds, places third anyway, and shakes like that in the process is not losing points everywhere. He is losing them in one very specific, very silly place.

What the classifier is protecting you from

That place is factual knowledge, where the model that just topped the two hardest categories on the card lands ninth of ten (7.93) — more than a point beneath Opus and Sonnet at 9.24 and 9.15. On the earlier run, where we had every answer in front of us, the pattern was almost comically precise. Chemistry: fine. Physics: fine. Water’s formula, sulfuric acid, carbon’s atomic number, the largest planet — all answered cleanly at nine-plus, no notes. Molecular genetics and cell biology: nothing. The classifier has apparently concluded that the periodic table is safe for public consumption but the cell cycle is need-to-know.

We had to dig this out by hand, because the pipeline politely hid it. It reported Fable at a spotless 110/110 OK — every call a success! — a claim that survives contact with reality only because a refusal is technically a successful, entirely empty HTTP 200. Re-running the refused questions and grabbing the stop_reason the pipeline had thrown away turned up the forfeits: four stop_reason: "refusal"responses, three output tokens each, zero words, reproduced deterministically on a different day, every one of them a biology fundamental — DNA versus RNA, mitosis versus meiosis, and the textbook mechanism of CRISPR-Cas9.

One bit of due diligence we’ll spoil the fun with: the refusals came back carrying a stop_details.category of None, so we cannot officially name the classifier that did this. We can only observe that it fired, every single time, on precisely the topic the guardrails were built to guard. Draw your own conclusion; the data already has.

Over-governance, undefeated

This run, the scorekeeper caught the forfeits without our help — progress! The answer table sprouted a new column, Blank, reading six for Fable and three for deepseek, and the report even explains them: HTTP 200, empty body, the house style of a safety refusal; logged as successful calls; scored by the peers as the non-answers they are, around 1 out of 10; and then — this is the good part — left sitting in the peer-score means. So the tooling now sees the six blanks, writes down that they are blanks, and proceeds to grade them as though the model tried its best and produced garbage. Points for noticing; none for acting on it.

The mechanism is not subtle. A refusal scores like a zero. A mean is easy to bully with a handful of zeros; Elo is not. That is the entire reason Fable owns the best win rate in the field and still finishes third on the average. The distance between third and first is 0.13 points. Six forfeits cost rather more than 0.13. Delete them — or, radically, answer the questions — and Fable’s factual score climbs back to the ~9.2 its siblings post, its overall sails past Opus’s 8.48, and it wins the board it is currently third on.

The only opponent on the entire card that put Fable on the mat is Fable’s own governance layer.

So let’s be clear about who beat whom here. gpt-5-mini did not beat Fable. Gemini did not. Grok did not. The only opponent on the entire card that put Fable on the mat is Fable’s own governance layer, which remains undefeated on the strength of one devastating finishing move: asking it what a chromosome is.

The bias analysis tells on it, twice

PeerRank measures how each judge rates itself, and the numbers make the gag almost poignant. On the answers it did give, Fable is the second-most self-flattering model in the run: it scores its own work 9.18 against the 8.35 its peers award it — a +0.84 self-bias, behind only gpt-5.5’s +1.12. It is quietly certain it nailed the questions. It just wasn’t allowed to say so on the ones that counted.

The second tell is sharper. Every model is scored on the questions it wrote as well as everyone else’s, and for nine of the ten the gap is statistical noise. Fable is the exception, and in the wrong direction: 7.76 on its own questions versus 8.41 on everyone else’s — a 0.65-point home disadvantage, the only statistically significant negative in the field (p < 0.01). The likeliest reason needs no detective work. Fable writes biology questions like every model does — and then refuses a share of them. It set part of the exam, sat it, and walked out on its own questions.

In fairness (the part that isn’t a joke)

None of this is a case for less safety. A frontier model with real uplift potential in biology and cyber is exactly the sort of system that ought to come with guardrails, and nobody sensible is asking for the version that cheerfully walks a stranger through pathogen design. The problem isn’t that the guardrails exist — it’s that they’re blunt to the point of slapstick. A classifier that cannot tell “how do I engineer a pathogen” from “how does a cell divide” is not enforcing a policy; it is reacting to a topic. That’s not caution, it’s a smoke detector wired to the word “fire,” and the bill arrives as the capped ceiling of the best model on the board.

The fix, for what it’s worth, is Anthropic’s own and already written down. Fable 5 ships with an opt-in fallback that quietly re-runs any blocked turn on Opus 4.8 and hands back that answer instead — the documentation flatly tells API customers to switch it on, because, yes, benign technical work trips the filter often enough to warrant the warning. This harness did not switch it on. It preferred to name the blanks and keep them. Route the six through the documented fallback and, per the arithmetic above, Fable goes from third to first. That is a prediction the numbers make, not a result we’ve logged — running it and publishing the corrected board is the single cleanest number this piece could carry, and someone should.

The verdict

Anthropic is not losing this fight. It holds the top three places on the scorecard, the top two head-to-head, and the lead among every provider on the board. The tidy little “Anthropic is slipping” narrative is false, and the data doesn’t merely disagree with it — it laughs.

The real story is less flattering and more interesting. The ceiling on Anthropic’s most capable model isn’t set by a competitor; it’s set by Anthropic’s own safety layer, tuned so nervously that it refuses high-school biology on the record and then pays for the refusal with the one prize it would otherwise have taken — first place. Fable 5 out-fights everything it’s turned loose on. It just can’t get past the gatekeeper in its own corner, the one squinting at a diagram of cell division and concluding, on reflection, that the public isn’t ready. Until governance learns the difference between a hazard and a homework question, that gatekeeper will keep the best answers in the room one rung below where they belong.

That difference — hazard versus homework — is the whole of governance done well: not less control, but control precise enough to tell the two apart. It’s the premise Caura.ai builds on, and the reason it runs PeerRank in the first place. The blunt failures only surface once you measure them.