Anthropic and Adaptyv Bio launch contest to lab-test 5,000 AI-designed proteins

Anthropic and Adaptyv Bio have launched a co-sponsored protein design competition that will put more than 5,000 AI-generated protein designs through experimental validation in automated wet labs. According to a dev.to report on the announcement, the program combines Claude-enabled biology models, laboratory testing, cloud compute and direct funding into a single end-to-end pipeline, and will concentrate on five selected protein-design challenges.

What participants receive

The support package is unusually complete for a competition of this kind. Citing Anthropic's update on Claude and biomolecular modeling, dev.to reports that eligible teams can receive:

• Up to $1 million in Claude credits for AI-assisted design work.
• Additional funding for experimental validation through Adaptyv Bio.
• Up to $250,000 in Modal compute credits for computational work.
• DNA synthesis provided by Twist Bioscience.
• Access to optimized biology models that will be open sourced after the initiative.

Access is structured rather than automatic: prospective participants must apply, and Anthropic frames the competition as part of its broader Life Sciences Verification Program.

Building on earlier results

The competition extends prior collaboration between the two companies. In an August 2026 case study, Adaptyv Bio tested Claude-driven protein design campaigns in its automated wet lab and reported 354 binders from 1,320 designs, a hit rate of roughly 26.8%. The dev.to report says binders were found for 14 targets, though it describes the campaign as covering 16 targets in one passage and "14 of 15" in another.

Those figures are company-reported results from an earlier campaign, not outcomes from the new competition, whose results have not yet been reported. They do explain the partners' reasoning for scaling up: the new program aims to validate roughly four times as many designs as the earlier case study while focusing them on five difficult problems.

Closing the design-test loop

The notable part of the announcement is not model access alone but the closed loop it creates. Designs can be proposed computationally with Claude, synthesized as DNA by Twist Bioscience, and measured in Adaptyv's lab. Protein engineering ultimately succeeds or fails on experimental evidence, and wet-lab validation is what exposes the gap between a promising computational prediction and a molecule that actually works. At a scale of more than 5,000 designs, the program should produce a far more informative picture of how well the workflow transfers across problems than any announcement could.

For smaller biotech teams and academic labs, the package addresses the stages that typically stall AI-enabled biology projects: large-scale model usage, synthesis capacity and lab iteration. A faster design-test-learn cycle could let such teams evaluate far more hypotheses than a sequential manual workflow, and the open-sourcing commitment could extend the resulting models beyond the original participants.

Open questions

The published information leaves several things unspecified, as dev.to notes: participant selection criteria, timelines, ownership terms for competition outputs, and the final evaluation metrics for each challenge. A credit package is also not a finished drug-discovery or protein-engineering platform. Domain expertise, careful experimental design and biological validation remain essential regardless of how much compute is on offer.

Why it matters

The competition is a concrete test of general-purpose AI inside a measurable scientific workflow. Rather than positioning Claude as a replacement for researchers or laboratory infrastructure, Anthropic and Adaptyv are embedding it in a pipeline where its outputs are checked against laboratory evidence, making the eventual results the real measure of the effort. If hit rates similar to the earlier campaign hold at the new scale, it strengthens the case for AI-assisted protein design as a practical engineering tool; if they do not, the distance between computational suggestion and laboratory outcome becomes the story. Either way, thousands of experimentally validated designs will add genuine evidence to a field where claims often rest on computation alone, and the open-sourced models could widen access to tools that smaller labs rarely get to build themselves.