Chemists confirm what a reaction made mainly by reading its nuclear magnetic resonance (NMR) spectrum, a list of peaks that has to be matched to the atoms of a proposed structure. This post, published on Anthropic's Science blog on 5 June 2026, asks whether a general-purpose language model can do that work as well as the dedicated software chemists use, and whether it can also run the problem backward, from spectrum to structure. The post introduces Anthropic's chemistry effort and summarizes a white paper by David Kamber, an Anthropic chemist, titled How Claude performs on NMR prediction and structure elucidation (13 pages, dated 5 June 2026, linked from the post as the "Full version"). The white paper is the version of record for the numbers; the post supplies the framing and the plans. Neither has been peer-reviewed, and the white paper does not describe its prompts.
What the study does
Two directions of NMR work
NMR spectra report one peak per chemically distinct hydrogen (¹H) or carbon (¹³C) atom. A peak's chemical shift, its position in parts per million (ppm), reflects the atom's chemical surroundings. A hydrogen peak is also split into a pattern, its multiplicity (singlet, doublet, triplet and so on), set by the number of neighboring hydrogens, and the spacing of the split sub-peaks is the J-coupling constant, measured in hertz (white paper, Introduction and its footnotes).
The white paper tests two tasks:
- Forward prediction: given a structure, predict its spectrum. ChemDraw and MestReNova, which the paper describes as on "essentially every chemist's desktop", do this; the versions tested were ChemDraw 25.0.2 and MestReNova 17.0.0 (title page).
- Inverse prediction, or structure elucidation: given a spectrum and a molecular formula, propose the structure. The paper says ChemDraw cannot do this, and MestReNova assigns peaks to a known structure but does not generate candidates (Introduction). So the inverse task has no software baseline.
The compounds
A chemist drew compounds from four ChemRxiv synthetic-chemistry preprints, read each one, kept compounds with complete and self-consistent NMR and high-resolution mass spectrometry (HRMS) data, and transcribed the peak lists by hand (white paper, Evaluation and "Selection protocol"). The rule was to take the first fully characterized novel compounds in each preprint, excluding compounds reported as mixtures of rotamers. The paper says "we chose and locked the compounds" before generating any predictions, to avoid selection bias. The post adds that the preprints were posted after the models' training cutoff "so as to avoid selection bias"; the white paper does not state this, and neither document gives the cutoff dates. Crossref's records date three of the preprints 21, 22 and 27 April 2026, and the fourth (Strong et al., on spirocyclization, which by its title is likely the source of the spiroketones; the pairing is our inference) 7 March 2025 (our check of the DOIs; the white paper's references give "2025" for it).
The forward set has 20 compounds in four scaffold classes of five each, chosen so that each class poses a different NMR difficulty: chloropyridazines (P1), Boc-N-aryl maleimides and N-Boc ynamides (P2), spiroketones (P3) and α-silyl methanesulfonamides (P4) (Figure 1). The inverse set of 15 was drawn separately under the same protocol.
How the forward task was scored
Each tool received the structure as a SMILES string and the solvent of the original report, and predicted ¹H and ¹³C shifts with multiplicities and J values. Each Claude model (Opus 4.7, Opus 4.6 and Sonnet 4.6) was queried three times per compound; the two classical tools are deterministic and were run once. Predicted peaks were matched one-to-one to experimental atoms by minimum error, and errors were computed against fixed totals of 401 hydrogen and 225 carbon atoms. ¹H multiplets reported as ranges wider than 0.3 ppm were dropped from the shift error but kept for multiplicity (white paper, "Forward prediction"). The headline tolerance is ±0.20 ppm for ¹H and ±1.0 ppm for ¹³C. A per-compound win rate counts the compounds on which a tool had the lowest mean absolute error (MAE).
How the inverse task was run
Opus 4.7 alone was given 15 problems, three times each, with the exact molecular formula from HRMS and the 1D ¹H and ¹³C spectra, and asked for up to three ranked SMILES candidates; stereochemistry was excluded because 1D NMR cannot fix it (white paper, "Inverse prediction"). The eight simpler targets got only those data. The seven structurally denser targets also got the SMILES of the reaction's starting material, "since accurate prediction on these scaffolds required that additional anchor." No reagents, conditions or mechanism were given in either condition. The paper does not say whether a run counted as correct when the target was the top candidate or anywhere among the three.
What it finds
Shift accuracy
All numbers are from the white paper, measured on the 20-compound forward set.
| Tool | ¹H MAE (ppm) | ¹H within ±0.20 ppm | ¹³C MAE (ppm) | ¹³C within ±1.0 ppm |
|---|---|---|---|---|
| Opus 4.7 | 0.079 | 91.2% | 1.375 | 65.9% |
| Opus 4.6 | 0.113 | 82.0% | 1.680 | 59.1% |
| Sonnet 4.6 | 0.145 | 82.3% | 2.152 | 49.5% |
| MestReNova | 0.150 | 52.4% | 1.484 | 58.2% |
| ChemDraw | 0.143 | 75.8% | 2.107 | 47.6% |
The MAE values are printed on Figure 2 and the tolerance shares on Figure 4; Claude values are means of three replicates. Opus 4.7 had the lowest ¹H error and, in the paper's word, was "comparable" to MestReNova on ¹³C (1.37 against 1.48 ppm). The post calls these two "effectively tied." Two conditions matter here:
- MestReNova's coverage. MestReNova predicted 267 of the 401 hydrogens. The paper notes that its ¹H MAE "is not directly comparable to fully covered tools" (section "Forward prediction"), Its low ¹H tolerance share (52.4%) appears to count the uncovered atoms as misses, since the paper states fixed denominators of 401 and 225 (our reading; 267 of 401 would cap it at 67%).
- Win rates. Opus 4.7 had the lowest per-compound error on a mean of 11.7 of 20 compounds for ¹H and 8.0 of 20 for ¹³C, the largest count on each nucleus (Figure 4, bottom). The text says Opus 4.7 "wins on most compounds on both nuclei"; on ¹³C, 8.0 of 20 is the most of any tool but fewer than half. MestReNova won 5.3 of 20 on ¹³C. The paper also notes that the winner on each compound shifts between replicates often enough that "single-replicate assessments would substantially misstate win rates."
Where the models erred
One hydrogen in the chloropyridazine family, an NH proton that exchanges slowly with the solvent, appears experimentally between 6.8 and 7.9 ppm. Opus 4.7 placed it somewhat low, Opus 4.6 scattered it across several ppm, and Sonnet 4.6 put it at 10 to 13 ppm (white paper, "Forward prediction"; Figure 5). Both Opus models predicted that family's aromatic ring carbons systematically low. Every tool, Claude and classical alike, predicted carbonyl carbons slightly low, so, the paper notes, a check that requires two tools to agree on a carbonyl shift will not catch the error.
Multiplicity and coupling constants
On ¹H multiplicity, scored over all 401 hydrogens, the exact-match shares were 85% for Opus 4.7, 83% for Sonnet 4.6, 78% for Opus 4.6, 54% for MestReNova (with 33% missing) and 41% for ChemDraw (Figure 7). The paper says ChemDraw's low share is "a scoring artifact rather than a capability gap": it commits to specific patterns where the experimentalist reported only an unresolved multiplet, and 42% of its calls fall in that lenient category.
On J values, 80% to 84% of the Claude models' matched pairs fell within ±0.5 Hz of experiment, against 26% to 35% for the classical tools (Figure 8). The sample is small: there are 33 anchored experimental J values across the 20 compounds, matched by 79 to 90 pooled predictions for each Claude model and by 23 (MestReNova) and 31 (ChemDraw) for the classical tools. The paper attributes part of ChemDraw's error to template defaults, such as a 12.4 Hz coupling in 5 of its 31 calls.
Stability across runs
The paper reports Opus 4.7 as the most stable model: its run-to-run variation in MAE is smaller than its margin over the next tool (Figure 2, which plots each replicate).
Structure elucidation
From formula and spectra alone, Opus 4.7 recovered all eight simpler targets on all three attempts. With the starting-material hint, it recovered four of the seven denser targets on all three attempts and the other three on two of three (white paper, "Inverse prediction"; Figure 9). That is 42 of 45 runs, our arithmetic. The post's limitations section says what happened without the hint: on the densest targets "the model could loop through its reasoning without committing to a final structure." The white paper reports no scores for that condition.
The authors' conclusion
The white paper concludes that "for routine data prediction Opus 4.7 is now as good as or better than ChemDraw and MestReNova on average" (Summary), and that the model "makes inverse elucidation from 1D NMR + HRMS tractable" (Takeaway). It contrasts this with dedicated elucidation software, which it says typically needs 2D NMR, specialized training and licensed tools. The post describes its overall claim as "a modest one": that Claude "is starting to meaningfully assist chemists" with translation, recall and integration work.
The prompts and inputs used
Neither document prints the prompts given to the models or says whether extended thinking or tools were used. The white paper specifies only the inputs: a SMILES string and solvent for the forward task, and formula, spectra and, for seven problems, a starting-material SMILES for the inverse task. The hint was assigned per problem by structural complexity, on the paper's statement that these scaffolds "required" it, so the two inverse conditions are not a controlled comparison.
Limitations and open questions
The white paper and the post list these limitations themselves:
- Small sample. 20 compounds in four scaffolds and 15 inverse problems; the rankings "should be read as indicative rather than precise" (white paper, Limitations). The post would want several hundred compounds across 20 to 30 scaffold classes, with at least 15 per class.
- Narrow chemistry. Slow-exchange NH heteroaromatics are represented only by chloropyridazines; 2D experiments and stereochemistry are out of scope by design, so complex natural products were not tested; only three solvents (DMSO-d₆, CDCl₃, D₂O) appear.
- The hint. The harder inverse problems were posed with the starting material because the model otherwise did not commit to an answer (post, Limitations).
Further points a reader may need:
- Who checked the results. The ground truth is the experimental peak lists published in the four preprints, transcribed by a chemist; the scoring is the author's. No outside chemist, journal or replication is reported, and the code and data for the comparison are not released.
- Unequal comparison. Claude models were scored as the mean of three runs and the classical tools on one deterministic run, and MestReNova's ¹H score covers two-thirds of the atoms.
- Training-data overlap. The post's statement that the preprints postdate the training cutoff is not detailed in either document, and one of the four preprints was posted in March 2025 (Crossref), more than a year before the post. Recency guards against contamination; it does not address selection bias, which the white paper handles by locking the compounds in advance.
- Developer's evaluation. The models and the evaluation both come from Anthropic, and the post introduces a program to make Claude "better at chemistry."
Research questions
- Do the rankings hold at the scale the post proposes? Repeating the forward task on several hundred compounds across 20 to 30 scaffold classes, as the post suggests, would show whether Opus 4.7's ¹H lead and ¹³C parity survive within-class variance.
- How often is the correct structure the top candidate? The inverse task allowed up to three ranked candidates. Reporting top-1 and top-3 recovery separately would show how much a chemist must still choose among answers.
- How does elucidation fail without the starting-material hint? Running the seven denser problems without the hint, and recording non-answers and wrong structures, would measure what the hint contributes.
- Is there contamination from the preprints? Comparing performance on compounds from preprints posted before and after each model's training cutoff would test whether recency affects accuracy.
- Do the J-coupling results hold with more data? The comparison rests on 33 experimental J values. A larger set, with matched pair counts reported per tool, would show whether the gap to classical tools persists.
- Would a general model and a classical tool together catch more errors? The shared carbonyl bias suggests some errors are common to all tools. Measuring disagreement-based error detection across tool pairs would show which combinations help.
How it connects
- Claude accelerates protein design and analytical chemistry (August 2026) cites this post (its footnote 9) as earlier work on NMR analysis, and moves from peak lists to raw instrument files processed by Claude Opus 5.
- The Claude Opus 4.7 system card describes the model that led most measures here; the link is this note's, not the post's.
- Introducing our Science Blog set out the blog this post appears in, and BioMysteryBench is another Science-blog evaluation of Claude on a domain task; both connections are this note's, not the post's.
References
- Anthropic. (2026, June 5). Making Claude a chemist. https://www.anthropic.com/research/making-claude-a-chemist record
- Kamber, D. (2026, June 5). How Claude performs on NMR prediction and structure elucidation: Assessing Claude Opus 4.7 against ChemDraw 25.0.2 and MestReNova 17.0.0 (white paper, 13 pp.). Anthropic. https://www-cdn.anthropic.com/07441e654ad3dfeb0cd090e9361511562825d012.pdf record
- The four ChemRxiv preprints the compounds came from, as listed in the white paper's references (not read for this note): https://doi.org/10.26434/chemrxiv.15002316/v1, https://doi.org/10.26434/chemrxiv.15002423/v1, https://doi.org/10.26434/chemrxiv-2025-59lfh, https://doi.org/10.26434/chemrxiv.15002274/v1