Anthropic has published its first major chemistry-focused research as part of an expanded science initiative, demonstrating that its general-purpose frontier model, Claude Opus 4.7, can match or exceed specialized desktop software that chemists have relied on for decades.
Published on June 5, 2026, in a research report titled “Making Claude a chemist,” the findings reveal that Claude can perform complex Nuclear Magnetic Resonance (NMR) spectroscopy tasks at a level competitive with industry-standard deterministic simulators and analytical engines like ChemDraw and MestReNova.
The Core Breakthrough: General AI vs. Specialized Software
NMR spectroscopy is an essential, time-consuming step in synthetic chemistry. Chemists use it to analyze molecular “fingerprints” to confirm exactly what molecule they have synthesized in a flask. Traditionally, this requires specialized software or manual parsing.
Anthropic tested Claude Opus 4.7 across 20 complex compounds sourced from recent synthetic chemistry preprints, evaluating the model on two distinct processing directions:
1. Forward Prediction (Simulating a Spectrum)
Given a drawn molecular structure, the AI simulated what its NMR spectrum should look like.
- Hydrogen ($\text{H}$) NMR Shifts: Claude Opus 4.7 posted the lowest average error in the study at $\pm 0.079 \text{ ppm}$. In chemistry, errors under $0.1 \text{ ppm}$ represent exceptionally high-quality structural predictions.
- Carbon ($\text{C}$) Shifts: Claude tied for first place with the dedicated software MestReNova, recording an average error of $\pm 1.37 \text{ ppm}$.
- Structural Nuances: The model outperformed traditional tools by a landslide on consistency metrics when predicting peak splitting patterns and J-coupling values.
2. Inverse Prediction (Structure Elucidation)
Going the other way, the model was given raw peak lists and high-resolution mass spectrometry data and tasked with working backward to deduce the hidden molecular structure.
- The Result: Opus 4.7 successfully recovered all simpler target structures on every attempt.
- Complex Structures: When provided with minor hints regarding starting materials, it successfully deduced four out of seven highly complex, dense molecular matrices.
Why This Upsets the Standard AI Benchmark Trend
What makes Anthropic’s breakthrough unusual is that Claude Opus 4.7 was not fine-tuned on chemistry-specific data to achieve these results.
[Legacy Software Workflow] ──► Requires Proprietary Licenses + 2D NMR Data Arrays
[Claude Opus 4.7 Workflow] ──► Accepts Raw, Copied Peak Lists + Vision-Parsed Sketches
Instead of requiring multi-dimensional, slow-to-run 2D NMR data—which has long been considered essential for complex structure mapping—Claude can reason through plain-text experimental data and published research papers. A chemist can simply copy and paste a raw, unformatted readout directly into a standard Claude chat window and receive a highly accurate structural proposal back, completely bypassing the need for an expensive, proprietary software license.
Anthropic’s Expanded Science Horizon
This chemistry upgrade is part of a larger, coordinated push by Anthropic to transition large language models from reading text to actively executing advanced scientific cognition.
Alongside the research paper, the lab launched its dedicated Anthropic Science platform, a repository focusing on practical workflows for scientists in physics, biology, and chemistry. Anthropic is also actively deploying API credits to high-impact external research labs and acts as a central partner in the multi-billion-dollar Genesis Mission—a joint initiative across industry and academia designed to use frontier AI to dramatically compress the timeline of American scientific discovery.