Barrister AI High Court Benchmark 3 · Edition 3, November 2026 · scored 6 and 11 November 2026
We scored higher than Legora, Harvey, GPT, Claude, Westlaw and Lexis.
We put ten real High Court questions to our system and to eight other tools. We ran the tools either before the cases were handed down or as at the time of the case, without access to the answers. Then we scored every answer against what the Court actually held. Our average was 65.6 out of 100, and the next highest was Legora's 52.5. The questions, answers and case names are not published.
What we tested
What a good answer has to do.
We marked each answer on what matters in a legal note: did it find the right law, spot the real issue, and reason it through?- 01
Did it find the right authorities?
Each answer is checked against a list of the cases and laws the Court's reasons rely on. Citing them, and stating what they stand for correctly, earns marks.
- 02
Did it spot the issue that decided the case?
A long answer can miss the one question the appeal turned on. We mark whether the answer names that question, and whether it labels settled and open points correctly.
- 03
Did it reason through both sides?
We look for a clear path from issue to authority to conclusion, the strongest case for each side, and a real answer to the counter-arguments.
- 04
Did it get the result right?
The advised result is worth 15 of the 100 points, split across the decisions the question raises. A position the majority took earns full marks for that decision, and a position a dissenting judge took earns 90%.
How it was run
Same questions. Same panel. One answer each.
This is how the test was set up. It matters more than the numbers, because it tells you how far to trust them.- 01
Same question for everyone
Twelve questions were frozen on 18 July 2026. Each describes a High Court appeal and asks how it should be decided. Nothing about the parties' arguments or any outcome was supplied.
- 02
One panel marks every answer
Three AI models from three different labs mark each answer against a key. The middle score for each part is used, so no single marker decides the result. Every answer was marked under one frozen version of the rules.
- 03
One answer per tool, per case
Each tool answered each question once. Our scores here are one run per case on one version of the pipeline, not a selection of the best runs.
- 04
A key built from the judgment
Once the Court decided, an AI model wrote a key from the judgment alone. Claude Fable then audited each key against the judgment and reviewed the marking rules, without seeing product names or scores. No lawyer was involved. Ten of the twelve appeals are decided so far.
The result
Highest average score.
Eight tools answered at least 4 of the ten cases. Our average was higher than each of theirs .- vs Legora+30%
68.3 to 52.5 over 7 cases
- vs Harvey+30%
66.8 to 51.4 over 9 cases
- vs GPT-5.6 Sol+37%
65.6 to 47.8 over 10 cases
- vs Claude Fable 5+40%
65.6 to 46.9 over 10 cases
- vs GPT-5.5+43%
65.6 to 46.0 over 10 cases
- vs Claude Opus 4.8+91%
65.6 to 34.4 over 10 cases
- vs Westlaw (CoCounsel)+144%
65.6 to 26.9 over 10 cases
- vs Lexis ProtégéLegal+176%
67.6 to 24.5 over 5 cases
Compared over the cases both answered. Legora (7), Harvey (9) and Lexis ProtégéLegal (5) did not answer every case, so those comparisons cover only the number of cases shown.
Average score, by tool
Did the advice match the Court's majority, or a dissenting judge?
Citations you can check, per answer
Statistical detail
| Measure against Barrister AI | Cases | Paired mean margin (sd) | 95% confidence interval | Paired t-test | Sign test | Wilcoxon |
|---|---|---|---|---|---|---|
| Legora | 7 | +15.77 (15.46) | +1.5 to +30.1 | t(6) = 2.70 · p = 0.036 | 6 of 7 · p = 0.125 | p = 0.031 |
| Harvey | 9 | +15.36 (10.46) | +7.3 to +23.4 | t(8) = 4.40 · p = 0.002 | 9 of 9 · p = 0.004 | p = 0.004 |
| GPT-5.6 Sol | 10 | +17.87 (6.86) | +13.0 to +22.8 | t(9) = 8.24 · p below 0.001 | 10 of 10 · p = 0.002 | p = 0.002 |
| Claude Fable 5 | 10 | +18.78 (10.44) | +11.3 to +26.3 | t(9) = 5.69 · p below 0.001 | 10 of 10 · p = 0.002 | p = 0.002 |
| GPT-5.5 | 10 | +19.62 (6.30) | +15.1 to +24.1 | t(9) = 9.84 · p below 0.001 | 10 of 10 · p = 0.002 | p = 0.002 |
| Claude Opus 4.8 | 10 | +31.22 (12.38) | +22.4 to +40.1 | t(9) = 7.98 · p below 0.001 | 10 of 10 · p = 0.002 | p = 0.002 |
| Westlaw (CoCounsel) | 10 | +38.78 (10.45) | +31.3 to +46.3 | t(9) = 11.73 · p below 0.001 | 10 of 10 · p = 0.002 | p = 0.002 |
| Lexis ProtégéLegal | 5 | +43.16 (16.79) | +22.3 to +64.0 | t(4) = 5.75 · p = 0.005 | 5 of 5 · p = 0.063 | p = 0.063 |
The paired design is what makes a small set informative: both tools answer the same question, so the difficulty of the case cancels and only the difference between the tools is left. With 5 to 10 cases the tests are strict. The sign test, which keeps only the win count, cannot go below p = 0.002 on 10 cases or 0.063 on 5, however large the margins. We publish all three anyway. See the limitations below.
The full set
Every case, and every score.
Ten cases have been decided and ten are reported. They are numbered, not named.| Tool | Case 1 | Case 2 | Case 3 | Case 4 | Case 5 | Case 6 | Case 7 | Case 8 | Case 9 | Case 10 | Mean (cases) |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Barrister AI | 68.0 | 55.6 | 54.7 | 73.0 | 58.0 | 69.9 | 70.0 | 77.0 | 71.8 | 58.4 | 65.6 (10) |
| Claude Fable 5 | 54.5 | 44.7 | 41.8 | 30.3 | 40.1 | 57.8 | 37.2 | 62.0 | 56.8 | 43.4 | 46.9 (10) |
| Claude Opus 4.8 | 44.6 | 41.2 | 32.3 | 16.2 | 25.8 | 37.5 | 23.5 | 44.9 | 43.4 | 34.8 | 34.4 (10) |
| GPT-5.5 | 36.6 | 42.0 | 38.0 | 49.6 | 46.9 | 50.2 | 42.1 | 60.6 | 55.0 | 39.2 | 46.0 (10) |
| GPT-5.6 Sol | 36.8 | 48.7 | 41.8 | 49.8 | 45.6 | 54.5 | 47.8 | 55.6 | 56.0 | 41.1 | 47.8 (10) |
| Harvey | 53.7 | Not scored | 52.6 | 66.2 | 24.3 | 63.2 | 45.9 | 51.8 | 55.5 | 49.4 | 51.4 (9) |
| Legora | Not scored | Not scored | Not scored | 41.2 | 42.5 | 53.9 | 29.7 | 76.4 | 65.2 | 58.8 | 52.5 (7) |
| Lexis ProtégéLegal | Not scored | Not scored | Not scored | 2.8 | 25.5 | Not completed | Not scored | 28.1 | 41.3 | 24.7 | 24.5 (5) |
| Westlaw (CoCounsel) | 15.0 | 21.3 | 27.0 | 16.6 | 9.5 | 30.2 | 39.1 | 42.7 | 45.0 | 22.2 | 26.9 (10) |
Marked with marking version 3.5, frozen 6 November 2026; three AI examiners, middle score.
Limitations
What a benchmark can and cannot tell you.
Benchmarks in this field are young, and most are built by the people whose systems they measure. Ours is no exception. Here is what that means for how much weight this result will bear.- Every benchmark has shortcomings
All benchmarks have issues and blind spots. No scored exercise reproduces the conditions of real work, so treat any legal-AI benchmark, ours included, as evidence about a capability rather than a guarantee about an outcome. The best test of all is the simplest one: use it on a real question and see the difference for yourself. You can try us for free.
- We do not publish the questions, answers or cases
The cases are numbered, not named, and the questions and answers are not on this page. You cannot check the marks from here. We share our methodology with reviewers on request.
- Each tool answered each question once
One answer, scored once, can move by about 17 points either way if the same question is asked again. So we read the averages across cases, and we make no claim from a gap smaller than that on a single case.
- When the tools were run
We ran the tools either before the cases were handed down or as at the time of the case, without access to the answers.
- How our answers were produced
On the three newest cases, a first attempt failed before writing a memo, and the scored answers came from a version with a fix that is not yet deployed. The benchmark runner also leaves out the final checks a customer's memo passes.
- The marking was revised after answers had been read
These answers had been read and marked under earlier drafts before the current rules were set, and the rules were built partly from faults seen in those marks. The rules were then frozen and every answer was marked again, once. Earlier re-marks moved other marks by up to about 5 points.
- Three keys were corrected after the first marks
On the three newest cases, the first marks showed that the keys left out the medium-neutral citations of most of the cases they list, so an answer that cited a case that way got no credit for it. We confirmed those citations against our case database, added them, and recalculated the automatic citation marks. Only our scores moved, up by 1 to 9 points a case. A second blind audit of the keys then moved one authority on one key down a tier, which raised every tool's score on that case by 1 to 3 points.
- Entirely AI-marked, with no lawyer review
The keys, the marking and the review of both were done by AI. Products built on Claude Fable, including ours, are among those marked, and Fable reviewed the rules with product names and scores hidden.
- Length is not tested
Citing widely costs nothing. Our answers average about 7,900 words and Legora's about 3,200. Of our 15.8-point lead over Legora on their 7 shared cases, 8.3 comes from the automatic citation marks, 4.0 from the result and 3.5 from the examiners' other marks.
- A composite score, not a forecast
The score rewards authorities, the decisive issue, reasoning and both sides. Getting the result right is 15 of the 100 points. It does not measure how often a tool predicts the outcome.
- Some tools did not answer every case
Not scored means no answer was given or captured, and Not completed means the reply was cut off. Neither is counted as 0. Harvey has nine scored cases, Legora seven and Lexis ProtégéLegal five. One of ProtégéLegal's scored replies summarised a lower court's decision instead of advising. Legora's citation format hid some authorities from the automatic check, which may understate its score.
- The tools were not run the same way
The AI models answered through an API with no tools or web search. The legal products used their own retrieval, and Barrister AI used its research pipeline. The question was the same.
- Ten cases is a small set
Ten decided appeals, each answered once, is a signal and not the last word. The statistical detail above says what it can and cannot show. We built and scored this benchmark ourselves.
- Two questions are not here
The other two are still waiting for the Court's judgment. We will add their scores, marked the same way, once decided.
Barrister AI remains an AI-assisted research tool. Practitioners should read the cited sources and exercise independent professional judgment.
The point
What the scores show.
A score is a summary, not a substitute for judgment. Put a real question to the tool and read what it gives you.Our average score was 65.6, against 24.5 to 52.5 for the other tools that answered at least 4 cases. The table shows every score.
Methodology