Anthropic Reports Claude Can Accelerate Protein Design and Chemistry Analysis

What Anthropic reported

On August 18, 2026, Anthropic published two experiments testing Claude on scientific workflows that normally require substantial specialist time. The first evaluated de novo protein-binder design; the second tested whether a generally available Claude model could process and interpret raw analytical-chemistry files.

Protein-binder design with wet-lab validation

Anthropic says Claude Opus 4.8 and Mythos Preview were used to design protein binders against 15 reported targets in an autonomous workflow that orchestrated specialist structure-design, sequence-design and co-folding models. External evaluators Adaptyv Bio and Twist Bioscience produced and tested the designs in the lab. Anthropic reports confirmed binders against 14 of the 15 reported targets, with overall hit rates of 22.6% for Opus 4.8 and 26.7% for Mythos Preview in a multi-target setup, and 35.1% for Mythos Preview in single-target runs.

The company also reports several high-affinity designs and cases where generated binders matched or exceeded previously reported results. These numbers are from Anthropic's study and should be interpreted in the context of its disclosed experimental setup, compute budget, target selection and external validation process rather than as a universal benchmark for automated drug discovery.

Claude on raw NMR and LC-MS data

In a separate analytical-chemistry test, Claude Opus 5 received raw NMR and LC-MS files plus short plain-language prompts. Anthropic reports that Claude returned processed results in 23 and 19 minutes respectively. Its purity estimate was 96.4%, compared with 96.33% from the contract laboratory, and its hydrogen-count analysis closely matched the lab's output.

The workflow is notable because Claude had to recover and interpret instrument data rather than simply summarize an existing report. Anthropic says the model also proposed a follow-up heavy-water experiment similar to one the lab independently performed and corrected one of its own initial overstatements after checking the data.

Why the result matters

The experiments suggest that frontier general-purpose models can increasingly act as orchestrators around specialist scientific software, reducing the manual coordination required for computational design and routine data analysis. That does not mean an AI model replaces wet-lab validation, domain expertise, safety review or downstream drug-development work. Protein binder discovery is only an early step in therapeutic development, and Anthropic explicitly describes advanced biological design as dual-use.

Availability and safeguards

Anthropic says Opus 5 remains generally available for analytical tasks, while some of its most capable life-science research capabilities remain restricted. The company says it is working toward a trusted-access program for scientists. The protein-design prompts, computational models and experimental data are being shared alongside technical reports for deeper scrutiny.