Astra for Law combines OpenAI's newest model with an index of US law spanning over 230 million URLs and instructions for legal analysis. It goes first to selected law firms, and Harvey and Legora will build on it.
- The model is the same; the score rises with the legal index, the instructions and the settings.
- The comparison with Claude Fable 5.1 is on an example OpenAI chose itself.
- The index is US law; for Bulgaria the direct benefit is zero for now.
In law a wrong answer is not a typo. You cite a ruling that was reversed by a higher court, and you lose the case, the client and your name.
OpenAI chose exactly this example for its own advertising, and against a competitor at that.
How to read the number
54 per cent sounds like half a job. On this test, though, the grading is strict, and what matters is the difference: the same model, with the same reasoning, goes from 38.7 to 54 with the legal index, the instructions and the settings. The model did not get smarter, it got a library and directions.
The comparison with Fable 5.1 is theirs, on an example they picked, and I do not take it as a verdict on the competitor. But the direction is right: a lawyer does not pay for pretty text, he pays for the citation to hold up before the judge.
For Bulgaria the direct benefit is zero for now; the index is US law. The interesting part will be who first does the same with our laws and case law, and whether it will be a company from outside or someone from here.