Beam 1.0 vs Mercury 3

The table below breaks down every workload axis where both models have published results. Each score is a normalized 0 to 1 average across the benchmarks tagged with that category. A gap of a few points is noise; a gap of ten or more is a real difference in capability. Where one model has not reported a result, the cell shows n/a and the missing benchmark does not drag its average down.

The headline benchmarks below are the most widely cited individual tests. GPQA Diamond measures graduate-level reasoning in physics, chemistry, and biology. SWE-Bench Verified tests whether a model can fix real GitHub issues. MMMU-Pro covers college-level multimodal understanding across six disciplines. AIME and FrontierMath push competitive and research math. BrowseComp measures web research ability.

Benchmarks tell you what a model can do in a controlled setting. They do not tell you whether it will work on your specific task, with your specific data, at your specific scale. The recommendations below map each common workload to whichever of these two models scores higher on the relevant axis. Use them as a starting point, not a final answer.