The argument for sending legal work to a frontier model has always been capability. Everything else — cost, latency, where the material goes — was a price worth paying because the alternative was worse answers. That argument was correct for a long time. It is worth checking whether it still is, because the answer decides whether running inference inside a firm's own boundary is an engineering choice or a sacrifice.

Two studies published in 2026 are useful here, and both are public.

Structured extraction: a domain-trained small model ahead of frontier baselines

In May 2026, a team at Onit AI Labs compared a self-hosted domain-trained model against five frontier systems on full-document structured extraction — 26 contract fields across six categories, evaluated on 24 held-out public SEC EDGAR contracts with 508 human-labelled field instances.[1]

The domain-trained model reached a macro F1 of 0.812 against 0.796 for the strongest frontier baseline, with micro F1 of 0.842 against 0.820. The cost difference was the larger result: roughly $0.018 per document batched, against $0.149 to $0.456 for the frontier models — a reduction of 78% to 97%.[1]

The authors are direct about the limits. Twenty-four contracts is a small evaluation set, they report no significance testing or inter-annotator agreement, the weakest fields were currency and dates, and amendments, statements of work and schedules were not covered.[1] This is a signal, not a settled finding.

Reasoning tasks: open-weight models at commercial parity, with prompting alone

A separate comparative study in March 2026 evaluated seven open-weight models under 10B parameters against commercial APIs on three legal benchmarks: contract entailment (ContractNLI), legal holding identification (CaseHOLD) and multi-label human-rights classification (ECtHR Task A).[2]

The best small model matched the commercial baseline on contract entailment — 77.5% against 76.6% — and beat it clearly on legal holdings, 71.2% against 62.2%. Mean accuracy across tasks was statistically indistinguishable from the commercial API. All of this was prompting only: no fine-tuning, which the authors describe as a lower bound on what these models can do.[2]

The failures are as informative. Multi-label classification collapsed for every model tested, small and large alike, with subset accuracy under 7%. A 9B model finished last overall, behind several smaller ones — size was a poor predictor of quality. And chain-of-thought prompting helped one task by 8.5 points while costing another 16.0.[2] There is no single configuration that is simply better.

What the evidence supports, and what it does not

Read together, these results support a narrower claim than "small models are as good now", and that narrower claim is the one that matters operationally.

  • Bounded, structured tasks — extraction, classification, entailment, coding a document set, building a chronology, producing a first draft against a template — are within reach of models a firm can run on its own hardware, at a fraction of the per-document cost.
  • Open-ended reasoning over novel material is a different claim, and neither study makes it. Nor does the harder end of the multi-label work, which nothing tested handled well.
  • Which model suits which task is an empirical question per task, not a procurement decision made once. That is an argument for routing across models under a policy, not for standardising on one.

Most of a matter's volume sits in the first category. That is what changes the economics: if the majority of the work can run locally and the minority that genuinely needs a frontier model can be routed to one under a policy the firm sets, then keeping material inside a controlled boundary stops costing capability.

Our own position, stated plainly

We train small language models for legal tasks, and we have not published benchmarks for them. Until we do, nothing above should be read as a claim about our models — it is a claim about what the field has demonstrated, from sources anyone can check.

When we publish, it will include the evaluation harness and the negative results. A benchmark without the cases the model failed is marketing, and this is an audience that can tell the difference.