Anthropic researchers, working with synthetic and computational chemists, are testing the effectiveness of Claude AI models in performing chemical analysis. The team evaluated three models—Opus 4.7, Opus 4.6, and Sonnet 4.6—on their ability to predict NMR spectra and deduce molecular structures, marking an effort to assist chemists with routine tasks like spectral analysis and translation between chemical representations. NMR spectroscopy (a method using magnetic fields to elucidate molecular structure) is a canonical but time-intensive step in synthetic chemistry. The team assessed the models against industry-standard software, ChemDraw and MestReNova, using 20 synthetic compounds pulled from ChemRxiv preprints published after the models' training cutoff to ensure unbiased results.

In forward prediction tasks, where tools predict the NMR spectrum from a known molecular structure, Opus 4.7 showed competitive accuracy. For hydrogen (¹H) spectra, Opus 4.7 achieved an average error of ±0.079 ppm, performing well within the tolerance window. On carbon (¹³C) prediction, Opus 4.7 and MestReNova were effectively tied with errors of ±1.37 ppm and ±1.48 ppm respectively. While classical tools like ChemDraw excelled in some areas, Claude models demonstrated superior performance in predicting peak splitting patterns and sub-peak spacing, achieving accuracy within half a hertz approximately 80% of the time, compared to 26 to 35% for the classical software. Claude also exhibited consistency across three independent test runs.

Beyond forward prediction, the researchers tested inverse prediction, or structure elucidation, where the model must identify a molecule from its experimental spectrum. Given 15 problems, Opus 4.7 correctly identified all eight simpler structures—single-ring or two-fragment molecules—using only their spectra and molecular formula. For seven more complex targets, such as fused rings and spirocycles, the model successfully identified the correct structure in all three test runs for four cases and in two of three runs for the remaining three, when provided with the starting material as a hint. The researchers noted that while these results indicate the models can assist with routine chemical analysis, limitations remain, including the small sample size, reliance on 1D NMR data rather than 2D experiments, and restricted solvent coverage. Anthropic plans to continue expanding Claude’s capabilities in synthetic reasoning, chemical literature understanding, and structure rendering to support ongoing scientific research.

Anthropic researchers, working with synthetic and computational chemists, are testing the effectiveness of Claude AI models in performing chemical analysis. The team evaluated three models—Opus 4.7, Opus 4.6, and Sonnet 4.6—on their ability to predict NMR spectra and deduce molecular structures, marking an effort to assist chemists with routine tasks like spectral analysis and translation between chemical representations. NMR spectroscopy (a method using magnetic fields to elucidate molecular structure) is a canonical but time-intensive step in synthetic chemistry. The team assessed the models against industry-standard software, ChemDraw and MestReNova, using 20 synthetic compounds pulled from ChemRxiv preprints published after the models' training cutoff to ensure unbiased results.

In forward prediction tasks, where tools predict the NMR spectrum from a known molecular structure, Opus 4.7 showed competitive accuracy. For hydrogen (¹H) spectra, Opus 4.7 achieved an average error of ±0.079 ppm, performing well within the tolerance window. On carbon (¹³C) prediction, Opus 4.7 and MestReNova were effectively tied with errors of ±1.37 ppm and ±1.48 ppm respectively. While classical tools like ChemDraw excelled in some areas, Claude models demonstrated superior performance in predicting peak splitting patterns and sub-peak spacing, achieving accuracy within half a hertz approximately 80% of the time, compared to 26 to 35% for the classical software. Claude also exhibited consistency across three independent test runs.

Beyond forward prediction, the researchers tested inverse prediction, or structure elucidation, where the model must identify a molecule from its experimental spectrum. Given 15 problems, Opus 4.7 correctly identified all eight simpler structures—single-ring or two-fragment molecules—using only their spectra and molecular formula. For seven more complex targets, such as fused rings and spirocycles, the model successfully identified the correct structure in all three test runs for four cases and in two of three runs for the remaining three, when provided with the starting material as a hint. The researchers noted that while these results indicate the models can assist with routine chemical analysis, limitations remain, including the small sample size, reliance on 1D NMR data rather than 2D experiments, and restricted solvent coverage. Anthropic plans to continue expanding Claude’s capabilities in synthetic reasoning, chemical literature understanding, and structure rendering to support ongoing scientific research.