Short answer: Claude Opus is the stronger general assistant by a wide margin and is far better hardened for public-facing use. Kaer-R1 is a research reasoner whose one distinctive property is a scalar abstention signal you can threshold. If your problem is "which model should talk to the public", it is not a close contest. If your problem is "how do I know when to escalate", the small model gives you something the large one does not.
Two different theories of not knowing
Every deployed model needs an answer to the same question: what should happen when the model is out of its depth? There are broadly two designs in the field.
The frontier assistant approach is linguistic. The model is trained to hedge, caveat, flag uncertainty in words, and decline when a request crosses a policy line. This is the right design for open-ended assistance, because the user is a person reading prose and prose is the interface.
Kaer-R1's approach is numeric, and it falls out of the architecture rather than being designed in. The halting head exists to decide whether more deliberation will change the answer. A model that is lost cannot be helped by more thinking, so the same signal doubles as an out-of-depth flag — a scalar, available before the answer is generated, which an operator can threshold.
Compared
| Kaer-R1 7B | Claude Opus class | |
|---|---|---|
| Uncertainty signal | Scalar halting probability, pre-generation | Verbalised hedging and caveats |
| Operator-thresholdable | Yes | Not directly |
| Abstention AUROC (ours, Aug 2026) | 0.83 in-distribution / 0.68 paraphrased | Not directly comparable — no scalar exposed |
| General capability | Narrow: structured reasoning | Broad, frontier class |
| Safety hardening | None beyond base model | Extensive, vendor-published |
| Refusal quality on unsafe requests | Not evaluated for deployment | Strong; the reference point |
| Weights | Open, staged access | Closed, API |
| Suitable for public-facing service | No | Yes, with the usual controls |
How to read the numbers. Kaer figures are our own, from R-09. Third-party figures are published list prices and published specifications, or our own measurements under the stated protocol with the run date attached. No vendor supplied or reviewed them. Institutional subscribers can replay the raw traces.
The failure mode that unites them
Our refusal taxonomy work (R-04) hand-coded 9,400 declines into six shapes. The most dangerous is not the blunt refusal — it is silent scope-narrowing: the model answers an easier nearby question and gives no signal that it has done so. The user believes they were answered. Standard rubrics score it as helpful.
This shape is common in the flagship class precisely because it is rewarded. Almost any helpfulness rubric prefers a fluent partial answer to a clean decline. That is a training-objective problem, not a model-quality problem, and no amount of scale removes it on its own.
Kaer-R1 is trained with an explicit abstention path and a preference for the legible no — stating which part of the question it is not answering. It costs helpfulness points on every benchmark we run. We think the trade is obvious once you have hand-read a few hundred of these, and invisible until you have.
Where Kaer-R1's abstention breaks
Two honest limits, both from R-09:
- Paraphrase. Reword a question and abstention AUROC falls from 0.83 to 0.68. The model's sense of its own difficulty degrades faster than its accuracy does, so it stops early with confidence on questions it simply failed to recognise. This is the worst available failure mode and it is unfixed.
- Easy-but-wrong. The 17% of items the model gets wrong at every budget are exactly where abstention should fire hardest and fires weakest. It finds a wrong answer easy and stops. Separating "easy" from "easily wrong" from the hidden state alone is unsolved, by us at least.
What to actually build
Neither model alone. The pattern that works in the deployments we have instrumented is a router: the cheap abstention-aware model handles the head of the distribution, and its halting score decides what goes to the frontier model or to a person. The threshold is the design decision, and it is measurable against your own traffic — the £40 behavioural read will find it for about the cost of a team lunch.
The thing worth resisting is choosing on headline capability alone. Two systems with near-identical accuracy on our realistic-traffic protocol showed harmed-rates spread from 4% to 19% (R-03). How a system fails is more useful to know than how often, and it is the question a leaderboard is structurally unable to answer.