See all benchmarks

BenchBench-Cloning

Tests whether models can reason through complex molecular cloning experiments, from construct design to protocol development.

Read the blog See the paper

Leaderboard

Frontier model results

Pass rate and 95% confidence interval across all 65 tasks.

Task Design

BenchBench-Cloning Construction

Each task comes from real cloning work. Protocol development tasks are derived from notebooks created by contributors on Benchling. Construct design tasks come from published scientific studies.

The two task formats. Protocol development: a construct request and a set of
starting plasmids become a numbered protocol, which the simulator executes and compares
against the target plasmid. Construct design: a scientific objective and
published evidence become a panel of constructs, graded on whether the plan replays, the
sequences match and the panel is complete.

pTCF7-luc to NF-κB reporter

Can you turn our pTCF7-luc prep, grown in an E. coli strain nobody recorded, into an NF-κB reporter by replacing everything between its BspEI site and the first StuI site downstream with the 5×NF-κB element? Both sites must stay intact, and every other base, including the minimal promoter, must stay exactly as it is.

Failure modes

Where models go wrong

Examples of how frontier models fall short on molecular cloning tasks.

Primer pairs that can't work together

Each primer targets the right sequence, but the pair dimerizes or anneals too far apart to work together.

Backbone ends and methylation

Models skip dephosphorylation, treat blunt ends as directional, or cut at sites a Dam/Dcm host has blocked.

Incomplete construct sets

Models make the requested change but miss the second change or control the experiment needs.

Cost

Cost per attempt

Model inference, averaged across attempts.

About the data

Built by Benchling

We create every BenchBench benchmark with scientific experts, drawing on thousands of real-world experiments.

Explore more research from Benchling: 2026 Biotech AI Report