Anthropic published results on August 18, 2026 showing that its large language models (LLMs) — Claude Opus 4.8 and the experimental Mythos Preview — autonomously completed a de novo protein binder design campaign with minimal human scientific intervention, achieving validated results that exceed standard industry benchmarks.
Experimental Design
<cite index="2-3,2-4">Anthropic described the exercise as an attempt to assess how much of the early-stage protein-design workflow could be handled by artificial intelligence (AI) with limited human intervention, instructing Claude Opus 4.8 and Mythos Preview to generate 30 candidate protein binders for each of 15 targets.</cite> <cite index="15-9">Claude carried out the work using publicly available specialist protein design models after receiving an initial prompt of approximately 30,000 tokens written by a human expert.</cite> <cite index="16-5,16-6">Human operators performed only three functions: approving network access, monitoring infrastructure, and forwarding sorted designs to the independent laboratories. Claude selected targets autonomously, chose binding epitopes, deployed tools, ran models, screened and optimized candidates, and called upon ten structure generation methods spanning 24 tool combinations.</cite>
Results
<cite index="17-9">External evaluators Adaptyv Bio and Twist Bioscience independently produced and tested Claude's designs in the lab, finding that of the 15 targets designed against, Claude successfully produced binders for 14 of them.</cite> <cite index="14-6">The test produced a total of 1,320 designs, of which 354 were confirmed by two independent laboratories to bind to target proteins, covering 14 targets.</cite>
<cite index="15-4">In a multi-arm design campaign, Mythos Preview and Opus 4.8 achieved overall hit rates of 26.7% and 22.6%, respectively, when designing against all targets simultaneously in a 48-hour session.</cite> <cite index="15-7">When Mythos Preview focused on individual targets in separate 24-hour sessions, its hit rate rose to 35.1%, compared with the 10% to 15% range Anthropic said is typical in protein design campaigns.</cite>
<cite index="3-4">Results include high-affinity binders against at least six targets, and binders matching or exceeding the best reported affinity against at least four targets.</cite> Against the target RBX1, a protein that drives targeted destruction of specific regulatory proteins inside cells, <cite index="4-7,4-8">Mythos Preview produced binders at a 40% hit rate. The same target had been the subject of an open design competition run by Adaptyv Bio, where the 245 entrants achieved a 3.7% hit rate.</cite>
<cite index="11-3">Clinically significant targets in the campaign included PD-L1, a checkpoint protein central to cancer immunotherapy; TREM2, implicated in Alzheimer's disease; TNFα, the target of blockbuster anti-inflammatory drugs including Humira; and EGFR, a well-established oncology target.</cite> <cite index="7-6">Anthropic noted that Opus 4.8 generated 12 effective designs against TNFα, with some candidates capable of binding human, cynomolgus monkey, and mouse versions of the protein simultaneously.</cite>
Validation and Methodology
<cite index="11-6">The research distinguishes itself from prior AI protein design claims in that Adaptyv Bio and Twist Bioscience physically synthesized and tested Claude's designs in wet-lab validation without modification.</cite> <cite index="5-6">Adaptyv Bio also published a separate case study detailing its benchmarking process.</cite> <cite index="8-10">Twist Bioscience received 1,260 of the 1,320 designs — those against every target except latent GDF-8 — and expressed them as human IgG1 Fc fusions in HEK293 cells.</cite>
Limitations and Safety Context
Anthropic was explicit about the boundaries of these findings. <cite index="6-7">The company clarified that "protein binders are not drugs" and that creating a high-affinity binder is only the first step in drug development.</cite> <cite index="14-12">The tests also revealed inconsistent performance across targets, and computationally generated binders still require structural validation, toxicology assessment, and clinical trials before they can approach drug candidacy.</cite> <cite index="6-10">Anthropic also highlighted the dual-use nature of its autonomous biological research capabilities, noting that it restricts protein design and other life-science research tasks in its most advanced model while developing safety measures.</cite>
<cite index="6-9">The company plans to launch an access program for scientists to utilize its advanced models, with further details expected soon.</cite>