Short answer: GPT-5 is more capable across more tasks. Kaer-R1 is roughly eleven times cheaper per solved task on hard reasoning, runs on your own hardware, and tells you when it is out of its depth. Those are different products, and the comparison only becomes useful once you say which problem you are solving.
This page exists because "which model is best" is the most-asked and least-answerable question in applied machine learning. Below is the version of it we can actually answer: for a defined workload, with a stated protocol, what does each system cost per correct answer, and how does each fail.
What each model is for
GPT-5 is a frontier general-purpose model. It is optimised for breadth — long context, tool use, multimodality, code, open-ended assistance — and it is served exclusively through an API. Kaer-R1 is a 7B dense open-weight reasoner with a learned halting head, optimised for one narrow thing: deciding, per question, how much deliberation a question deserves, and stopping when more thinking will not help.
Neither of those descriptions is a criticism of the other. A lab with our budget cannot build a frontier general model and should not pretend otherwise. What a small lab can do is take one axis seriously enough to move it.
Head to head
| Kaer-R1 7B | GPT-5 class | |
|---|---|---|
| Weights | Open, staged access | Closed, API only |
| Self-hostable | Yes — single GPU | No |
| Built for | Reliable reasoning at low cost | General capability, breadth |
| Compute allocation | Adaptive — learned per question | Fixed or operator-set |
| Hard-reasoning suite (ours, Aug 2026) | 61.4% | Higher — frontier class leads this suite |
| Cost / 1k suite items (measured) | £3.10 | Frontier list pricing, ~10–15× higher |
| Abstention signal | Built in (AUROC 0.83 in-distribution) | Verbalised confidence only |
| Evaluation card | Published, incl. failed variants | Vendor system card |
| Data residency | Your infrastructure | Vendor infrastructure |
How to read the numbers. Figures for Kaer models are our own, from the runs described in R-09. Figures for third-party models are (a) published list prices and published specifications, or (b) our own measurements under the Ordinary Tuesday and paraphrase protocols, labelled as such with the run date. They are not vendor-supplied and no vendor has reviewed them. Public Atlas releases blind model identities by class; where we name a model here we are naming the class it sits in, and the raw traces are available to institutional subscribers who want to replay the arithmetic.
Why cost per solved task is the honest metric
Benchmark tables almost always report accuracy and almost never report what the accuracy cost. This is how a model that samples sixty-four times and takes a majority vote ends up in the same column as a model that answered once. On our suite, Kaer-R1 with sixteen-sample voting scores 63.2% — better than its own single-pass 61.4% — at £236 per thousand items instead of £3.10. Same model, same weights, seventy-six times the bill, under two points of gain.
Frontier models are not exempt from this arithmetic; they are subject to a harsher version of it, because their per-token price is higher to begin with. The question a deploying team should ask is not "which model scores highest" but "at my volume, what does one additional correct answer cost me, and can I afford the tail".
Where the small model actually wins: knowing it is lost
The result from R-09 that we did not expect is the one that matters most in production. Kaer-R1's halting head has to estimate whether further deliberation will change the answer. A model that is lost cannot be helped by more thinking — so the same signal separates "I have this" from "I am out of my depth", before the answer is generated.
Used as an abstention score it reaches AUROC 0.83 in-distribution. Routing the bottom decile of traffic to a human or a larger model removes 41% of remaining errors while touching a tenth of the volume. For any service with a fallback path, that number is worth more than the capability gap, because it converts an unknown error rate into a budgeted one.
Frontier models express confidence in words, and the calibration literature is consistent that preference-tuned models are fluent in the register of confidence beyond its warrant. A verbalised "I'm fairly sure" is not a routable signal. A scalar you can threshold is.
Where Kaer-R1 loses, plainly
- Breadth. Anything outside structured reasoning — open-ended writing, multimodal work, long-context synthesis, agentic tool chains — is not what it was trained for and it shows.
- Paraphrase robustness of the halting head. Reword a question and abstention AUROC falls from 0.83 to 0.68. It under-thinks confidently. We have not fixed this.
- Safety hardening. Kaer-R1 is a research release with no hardening beyond its base model. Frontier models ship with far more work behind them and should be assumed safer in front of the public.
- Latency predictability. Per-item budgeting means per-item variance. A service with a hard latency ceiling should set a fixed budget and accept the lower score.
The decision, in one table
| If your constraint is… | Choose |
|---|---|
| Widest capability, fastest to ship | Frontier API model |
| Cost at volume on bounded reasoning tasks | Kaer-R1, with routing |
| Data cannot leave your infrastructure | Open weights — Kaer-R1 or another |
| You must show a regulator how it decided | Open weights, published eval card |
| You need a fallback path that fires reliably | Kaer-R1's abstention signal |
| Open-ended assistance for the public | Frontier model, hardened |
In most deployments we have instrumented the answer is both: a cheap abstention-aware model on the head of the distribution and a frontier model on the tail it flags. The routing threshold is the real design decision, and it is measurable on your own traffic in an afternoon. The £40 behavioural read is the cheapest way we know to find it.