| name | making-claude-chemist |
| description | Methodology from Anthropic research (Jun 2026) on benchmarking LLM capability for chemistry tasks, specifically NMR spectral analysis and molecular structure elucidation. Opus 4.7 achieves competitive accuracy with ChemDraw/MestReNova on hydrogen NMR (±0.079 ppm error), carbon NMR, peak shape prediction, and 1D inverse structure elucidation. Use when evaluating LLM performance for chemistry workflows, building AI-assisted molecular analysis tools, or understanding chemistry-specific AI benchmarks. Activation: Claude chemist, NMR spectral analysis, molecular structure elucidation, AI chemistry benchmark, Opus chemistry, Anthropic chemistry, chemical structure from spectra. |
| version | 1.0.0 |
| author | Anthropic Research |
| date | 2026-06-05T00:00:00.000Z |
| source | https://www.anthropic.com/research/making-claude-a-chemist |
| category | ai_collection |
| tags | ["chemistry","NMR","molecular-analysis","benchmark","spectral-analysis","structure-elucidation"] |
| activation_keywords | ["Claude chemist","NMR spectral analysis","molecular structure elucidation","AI chemistry benchmark","Opus chemistry","chemical NMR","Anthropic chemistry","1D inverse elucidation","chemical structure from spectra"] |
Making Claude a Chemist
Core Problem
Chemistry involves translating between multiple molecular representations:
- SMILES (text strings encoding molecular graphs)
- Chemical structures (2D/3D drawings)
- NMR spectra (instrument readouts showing peak positions)
- Molecular formulas (elemental composition)
- Journal figures (visual representations in papers)
Translating between these formats is time-consuming and error-prone. AI tools have been positioned as transformative for chemistry but remained largely aspirational until recently.
Key Finding
Opus 4.7 is competitive with professional chemistry software (ChemDraw, MestReNova) on NMR analysis tasks:
Hydrogen NMR (¹H)
- Opus 4.7 average error: ±0.079 ppm (well under half the tolerance window)
- Highest share of peaks landing inside tolerance vs. alternatives
- Evaluated across 3 runs per compound with averaging
Carbon NMR (¹³C)
- Opus 4.7 and MestReNova performed comparably
- ChemDraw showed wider gap
Peak Shape & Coupling
- Opus 4.7 less accurate on predicting peak shapes and coupling distances
- These features are also harder for human chemists
Structure Elucidation (1D Inverse)
- 8 simpler structures: 100% recovery rate from spectra + formula
- 7 harder targets: Correct structure recovered on all attempts (with starting-material hint)
- General-purpose LLM makes 1D inverse elucidation tractable
Methodology
1. Multi-Tool Comparison Framework
- Query LLM 3 times per compound, average results (accounts for output variance)
- Compare against deterministic tools (ChemDraw, MestReNova run once each)
- Use ppm tolerance windows for peak position validation
2. Task Decomposition
- Peak position prediction: Chemical shift values in ppm
- Peak shape prediction: Multiplicity (singlet, doublet, etc.)
- Coupling constant prediction: Peak separation distances
- Structure elucidation: Given NMR + formula, deduce molecular structure
3. Benchmarking Protocol
