How well AI models do the work lawyers actually do.
Lawve-bench-1 scores AI models on in-house legal tasks drawn from real practice: the drafting, review, research and multi-step matter work legal teams handle every day. Each model gets a Lawve Index, a single score out of 100. The higher it is, the better the model handled the work.
Best right now: Claude Opus 5.5 max leads with a Lawve Index of 63.
Most AI models let you choose how hard they think before answering, from low to max effort. More effort usually gives better answers, but each task then costs more and takes longer. In the chart below, each line is one model and each dot is one effort setting. Higher means a better score; further right means cheaper. A line that sits high and far to the right gives strong results without a large bill.
Take Claude Opus 5.5 low: it scores 52 for $0.27 a task. At max effort the same model reaches 63, but each task costs $4.11.
Overview
The next chart shows every model and effort setting we tested. The dashed line is the : the highest score you can get at each price. A dot below the line is outscored by a cheaper or equally priced option, so it is rarely the best pick. The cost axis grows tenfold at each step, which keeps very cheap and very expensive models readable on the same chart.
Cost
Speed matters when someone is waiting on the answer. This chart works the same way, with the average time a model takes per task instead of its cost. Dots on the dashed line are the fastest way to reach each score. Further left is faster. GPT-6.1 Sol low, for example, reaches 50 in about 30s a task.
Speed
Claude Opus 5.5 max, if budget allows. It leads Lawve-bench-1 with a Lawve Index of 63 at max effort. The right choice also depends on budget and turnaround: GPT-6.1 Sol xhigh reaches 57 to 58 for under $0.50 a task, and GPT-6 Astra max posts the strongest Non-Hallucination scores in the top ten.
Yes, this one. Lawve-bench-1 measures how AI models perform on real legal tasks across six areas: legal knowledge, agentic knowledge work, reasoning, long-context, non-hallucination and agentic tool use. The tasks are built in-house and kept private, so models cannot be trained on them.
Claude Opus 5.5 max, with 63. It is followed by Claude Opus 5.5 xhigh at 61. GPT-6 Astra max and Claude Opus 5.5 high share third place at 59. The Claude runs listed here use fallback.
To keep the scores honest. Public test sets end up in training data and reward models tuned to the test rather than to the work. Keeping ours private means a score reflects the work, not memorization.
As a starting point, not a verdict. It is a single score out of 100 that summarizes the six areas; higher is better. A gap of a point or two rarely matters in practice. Weigh it against cost and time per task, and always review AI output before relying on it.