Short answer: On bounded reasoning workloads in 2026 a well-chosen small model will get you most of the way to frontier quality for a small fraction of the cost — but only if you compare on cost per solved task, test with your own phrasings, and check whether the model exposes any usable signal about its own uncertainty. Most buyer's guides compare on none of those.
Why this category exists now
Two things changed. Inference-time deliberation turned out to buy real capability, so a small model given room to think competes on tasks it previously could not. And the cost of a competent answer fell far enough that the binding constraint on deployment moved from capability to price — the difference between a service that answers the questions it must and one that answers the questions people actually have.
The catch is that deliberation is not free, and most reported gains in this category are bought with a compute multiplier that the headline number quietly omits.
The six questions that decide it
- What does one correct answer cost? Not accuracy — cost per solved task. Divide what you paid by the number of items it got right. This single ratio collapses most of the marketing.
- How many samples produced that number? A score from a 64-sample majority vote and a score from one greedy pass are not the same measurement and should never share a column.
- Does it expose an uncertainty signal you can threshold? Verbalised hedging is not routable. A scalar is. If a model can tell you when to escalate, that is worth more than several points of accuracy to any service with a fallback path.
- What is the paraphrase spread? We found a 19-point median accuracy spread across twelve phrasings of the same question, against an 11-point gap between adjacent-tier models (R-07). A single-phrasing score is a poll of one person.
- How does it fail, not how often? Harmed-rates spread 4% to 19% across models with near-identical headline accuracy (R-03). Failure style is the part you design around.
- Can you host it? Licence, parameter count and quantisation decide whether this is an infrastructure decision or a procurement one.
What the numbers look like when you ask properly
From our own hard-reasoning suite — 4,200 items across competition mathematics, multi-hop retrieval, code repair and legal reasoning, single greedy pass, five seeds, cost measured rather than estimated:
| Configuration | Suite score | Mean thinking tokens | Cost / 1k items |
|---|---|---|---|
| 7B, fixed 4,096-token budget | 58.1% | 4,096 | £14.80 |
| 7B, adaptive halting | 61.4% | 790 | £3.10 |
| Reference 80B reasoner | 66.7% | 3,400 | £38.60 |
| 7B, 16-sample vote | 63.2% | 65,536 | £236.00 |
Three things to take from that table. Adaptive beat fixed-budget on the same weights while using a fifth of the tokens — more thinking is not monotonically better. The 80B still wins on score, at twelve times the cost of the adaptive 7B. And sixteen-sample voting is the worst deal on the page: under two points over adaptive, at seventy-six times the price.
The trap: benchmarks you can see are benchmarks that got trained on
Any public benchmark is a training target, and a model tuned toward one will look better than it is. This is not an accusation against anyone; it is a structural property of publishing a test. It is why our scenario banks stay private while the protocol, rubric and per-model results stay open, and why we run 13-gram overlap decontamination as a pre-registered step on every training mix rather than as a check when a number looks suspiciously good (R-08).
Practically: weight public-benchmark scores lightly, and weight a read on your own traffic heavily. The £40 behavioural read gives Kendall τ = 0.83 rank agreement with our full protocol at about 1% of the cost, which is cheap enough to run on every candidate rather than only the finalists.
Where small models are still the wrong answer
- Open-ended assistance for the public — use a hardened frontier model.
- Long-context synthesis, multimodality, agentic tool chains.
- Anything with a hard latency ceiling, if the model uses adaptive budgeting: per-item budgeting means per-item variance.
- Multilingual work outside the model's training distribution. Our own corpus is 89% English and we say so.
The pattern that works is not choosing. It is routing: a cheap abstention-aware model on the head of the distribution, a frontier model or a person on the tail it flags. In the deployments we have instrumented, 60 to 80% of volume never needed the expensive model at all — and the threshold that finds that split is measurable on your traffic in an afternoon.