New Delhi [India], July 14: For much of the past decade, progress in artificial intelligence has been measured by scale. General-purpose models grew larger and more capable, able to answer questions across almost every field. A different phase is now taking shape. In professions such as law, medicine, finance and engineering, the more relevant question is no longer which model knows the most about everything, but which is built to understand one field exceptionally well.
India’s All India Bar Examination has become an unexpectedly clear window onto that shift. In the most recent evaluation, LeXi AI, a platform built specifically for the Indian legal system, recorded the highest score among five AI systems tested on the examination, ahead of GPT 5.5, Gemini 3.1 Pro, Claude Opus 4.8 and DeepSeek v3.2. It is the second consecutive evaluation in which the domain-specific model has led the field.
A public benchmark, a repeatable result
The AIBE occupies a distinct position in the Indian legal system. Conducted by the Bar Council of India, it is the qualifying examination that every aspiring advocate must clear before entering practice. Both the paper and the official answer key are published, the benchmark can be independently reproduced by anyone who chooses to run it.
In this evaluation, all five models were given the official AIBE 20 paper under identical conditions, with no specialised prompting, and each answer was scored against the Bar Council’s published key.
The results were as follows.
LeXi AI answered 94 of 95 questions correctly, with one incorrect and none unanswered, for an accuracy of 98.95%.
GPT 5.5 followed with 92 correct answers, three incorrect and none unanswered, for an accuracy of 96.84%. Gemini 3.1 Pro answered 91 questions correctly, with three incorrect and one unanswered, for an accuracy of 95.79%. Claude Opus 4.8 recorded 86 correct answers, with nine incorrect and none unanswered, for an accuracy of 90.53%. Deepseek v3.2 answered 85 questions correctly, with ten incorrect and none unanswered, for an accuracy of 89.47%.
Why this matters
Artificial intelligence is moving quickly from experimentation into everyday professional work. Law firms are testing AI-assisted research and drafting, corporate legal teams are evaluating it for contract review and compliance, and courts in several jurisdictions are examining its role in improving access to justice. As that adoption deepens, one question matters more than any other: which systems can legal professionals rely on when accuracy is not optional.
The AIBE result does not answer every part of that question. What it provides is measurable, publicly verifiable evidence that a system designed specifically around Indian law can outperform leading general-purpose models on a transparent legal benchmark.
The rise of vertical AI
For years, the frontier of AI has been defined by general-purpose systems, and those systems have changed how people search, write and solve everyday problems. Professional work makes different demands. Law rests on statutes, judicial precedent, procedural rules and jurisdiction-specific reasoning that cannot be treated as general knowledge. On the evidence of benchmarks like this one, depth of domain expertise is coming to matter as much as the scale of the underlying model. Purpose-built systems approach the problem differently, training on a single field in depth rather than attempting to serve every profession at once.
More than a benchmark
The AIBE measures one dimension of legal capability. It does not test advocacy, judgment, ethics, negotiation or courtroom strategy, all of which remain human responsibilities. Benchmarks matter for a narrower reason, which is that they offer transparent and measurable evidence of technical capability. The significance here is not simply that one system scored highest. It is that a model built around Indian legal reasoning has now shown a measurable advantage over leading general-purpose systems on the country’s principal legal qualifying examination, across two consecutive evaluations.
Comment from LeXi AI
AI is entering a new phase. The future will not belong only to systems that know a little about everything. It will increasingly belong to those that understand one profession exceptionally well.
Law is among the most demanding of those professions. As a practising advocate myself, I do not believe lawyers need another general chatbot. They need technology that understands the language, structure and responsibilities of Indian law.
This result is encouraging because it supports that approach, and because it was measured on a public benchmark that anyone can check.
