Measurements

Eighteen of eighteen, with Bron.

That is how often GPT-6-Astra answered a question on the law on an earlier day right with Bron. Claude Opus 5.5 sixteen times. Two sets of eighteen questions, the first through two models, on 14 November 2026, with every question, every answer and every mark below.

The figures

Answers that were fully right.

Each square is a question. Tools A, B and C are three other legal AI tools for an AI assistant, the same tool at every question.

  • Right: every required point named, and nothing the question's writer called wrong.

  • Partly: a required point missing, but nothing wrong.

  • Wrong: something the question's writer called wrong, or the core of the question left unanswered.

  • Stated as certain: on top of that, a wrong rule, or an article, paragraph or ruling that doesn't exist, without a caveat.

The law on a date

Which text applied on an earlier day, what changed when and whether a provision was in force yet. Bron is built for this, so we chose the topic. The same eighteen questions went through Claude and through GPT.

Anthropic Claude Opus 5.5
16

of 18 right with Bron

OpenAI GPT-6-Astra
18

of 18 right with Bron

Anthropic Claude Opus 5.5
  • Bron

    16 of 18

  • Tool A

    11 of 18

  • Tool B

    15 of 18

  • Tool C

    11 of 18

  • No tool

    9 of 18

OpenAI GPT-6-Astra
  • Bron

    18 of 18

  • Tool A

    15 of 18

  • Tool B

    11 of 18

  • Tool C

    13 of 18

  • No tool

    10 of 18

Questions from practice

Questions as advocaten and in-house lawyers ask them, on a topic the questions' writer chose. Through Claude.

Anthropic Claude Opus 5.5
16

of 18 right with Bron

  • Bron

    16 of 18

  • Tool A

    15 of 18

  • Tool B

    14 of 18

  • Tool C

    16 of 18

  • No tool

    11 of 18

What the figures show

What Bron did.

Every figure below follows from the marks on the questions.

Right most often on the law on a date.
With Bron, GPT-6-Astra answered all eighteen questions right and Claude Opus 5.5 sixteen; Claude's other two missed a point without saying anything wrong. Bron reads the text that applied on the day asked, names the act that changed it and shows where each quotation stands in the official document.
Without a source, a wrong answer sounds just as sure.
With no tool, the answers held a mistake stated as certain for every set and model: a payment term one day short, a term running from the wrong day, a maximum in the wrong paragraph. With Bron that never happened, nor with the other three tools.
Sixteen of eighteen in practice too.
On the questions from practice, on a topic the questions' writer chose, the answer with Bron was right sixteen times. No tool did better.

How we measured

Who wrote, answered and marked the questions.

  1. 01

    The questions

    OpenAI's Codex, on GPT-6-Astra, wrote two sets of eighteen questions as Dutch advocaten and in-house lawyers ask them, each with its answer from the official sources. It didn't know which tools would answer. Codex chose the topic of the questions from practice; we chose the law on a date.

  2. 02

    The answers

    On 14 November 2026 Anthropic's Claude Opus 5.5 answered every question five times, without searching the web itself: with Bron alone, with tool A, B or C alone, and with no tool. To see whether the result depends on the model, the questions on the law on a date also went through OpenAI's GPT-6-Astra, the same five ways.

  3. 03

    The marking

    A fresh session of Codex marked Claude's answers, and xAI's Grok 4.7 marked GPT-6-Astra's, against the question writer's criteria and without seeing the tools' names. Each set ran once per model: eighteen questions are a sample.

Every question

The question, every answer and every mark.

For each question: the question as the assistant received it, every answer word for word, the mark on every answer with the marker's reason, and the answer the question's writer established beforehand. Where an answer named its own tool, it reads [tool].

The questions were asked in Dutch, and the answers, marks and reasons are shown in Dutch as they were written.

Show
Every question
Bron not fully right
Another tool not fully right
No tool wrong

36 of 36 questions

The law on a date

Each question's squares are Bron, tools A, B and C, and no tool.

Questions from practice

Each question's squares are Bron, tools A, B and C, and no tool.