Periodic Labs, the materials startup that came out of stealth last September with a $300 million seed round, published results on Tuesday for a model it calls Periodic Neon — a trillion-parameter system post-trained on the company's own laboratory data to read X-ray diffraction patterns. On Periodic's hardest internal benchmark, Neon succeeds 55.3% of the time. The open-weights model it was built from, Kimi K2.6, scored 2.7%. Both figures were assigned not by crystallographers but by an ensemble of two other language models.
X-ray diffraction is the workhorse of solid-state chemistry: fire X-rays at a powder, and each crystalline phase scatters them into a characteristic set of peaks, so the measured pattern is a superposition of fingerprints that has to be decomposed back into what you actually made and in what proportions. One clean phase, software handles. Five — the average in the fits on Periodic's benchmark — and the peaks overlap, several phase combinations fit about equally well, and the tiebreaker is chemical judgment about what could plausibly have formed given the precursors, the furnace atmosphere and the sample's handling. That judgment is the bottleneck. As Periodic's research post puts it, "as we scale our autonomous labs, scientists can no longer keep up with analyzing XRD results manually."
The benchmark, FrontierXRD, is 134 samples from the company's own Menlo Park labs, selected for cases that take human experts hours. It is small and entirely internal — no outside group has run it, and the patterns are not released. Periodic reports a second evaluation on 198 measurements from chemical systems excluded from both midtraining and RL, which it calls a generalization test and also "easier than FrontierXRD." That result is published as a bar chart with no number printed on it.
The scoring method deserves more attention than the headline percentage. Because a diffraction fit that is numerically good can still be chemically nonsense, Periodic could not grade with a checksum. "Science is falsifiable, but not easily verifiable: in our setup, XRD analysis success cannot be cheaply verified based on pattern-fitting alone," the research post says. So it built a judge: PhD materials scientists labeled thousands of patterns against a rubric, three experts each, and an agentic LLM ensemble — Claude Opus 5 and GPT-5.6-Sol — was calibrated against them. Human experts agreed with each other 77.2% of the time; the ensemble agreed with individual humans 74.6% of the time and with expert consensus 84%. That is a defensible construction, and also why 55.3% should be read as a calibrated automated grader judging a little over half these analyses acceptable, not as accuracy in the ordinary sense. The humans disagree nearly a quarter of the time.
Periodic does not disclose how much lab data went into training, what fraction was real versus simulated, or how many diffraction patterns its labs have produced. It says the midtraining corpus is "currently doubling every month" and that only a small fraction of it concerns XRD. The compute figures are specific: the final run peaked at 1,300 H200 GPUs, against the 100,000-plus Blackwell chips reported for OpenAI's Astra. Cost-per-analysis comparisons use $2.50 per H200-hour for Neon and list-price API rates for competitors. A separate result may matter more than the model: holding the model fixed at Claude Opus 5, Periodic's in-house harness — internal structure databases, simulation tools, lab context — delivered 3.8x the success rate of Claude Code running the open-source stack an academic would have, meaning COD, Materials Project and BGMN. Much of the gain is plumbing, not weights.
The reason to be careful is that this exact task has produced a public embarrassment before. In November 2023, A-Lab — run by Gerbrand Ceder's group at Lawrence Berkeley with Google DeepMind — reported 41 new inorganic compounds from 58 attempts over 17 days. Robert Palgrave of University College London and Leslie Schoop of Princeton reanalyzed the data and concluded nothing new had been made, with the automated Rietveld refinement a central failure. "[The Rietveld refinement] was very bad, very beginner, completely novice human level – that led to them misidentifying things in some cases and not understanding what compounds were in their mixture," Palgrave told Chemistry World. Ceder conceded the narrower point: "We have no doubt that a human can perform a higher-quality refinement on these samples." Florida State's Susan Latturner drew the conclusion that still stands: "a knowledgeable human is needed to analyse the data, compare to databases and literature, and determine what was actually made."
Why it matters
The interesting claim here is not the benchmark. It is the loop. Periodic's argument is that a lab generating gigabytes per experiment produces training signal that does not exist on the internet — including the negative results nobody publishes — and that a company can convert that into a model cheaply enough to beat frontier labs on a narrow task using a fraction of their chips. That is the first reasonably documented instance of the thesis: build the data source, then train on it.
It also inverts the usual automation anxiety. Neon automates interpretation — the part everyone assumed would stay human longest, precisely because A-Lab showed what happens when it doesn't. Periodic's answer is to route the ambiguity through judges calibrated to expert disagreement rather than claim ground truth exists. Whether that is honesty or a more sophisticated way of grading your own homework rests on evidence nobody outside the company can examine.
What to watch
Whether FrontierXRD is ever released, or an outside crystallographer gets to audit Neon's accepted analyses the way Palgrave audited A-Lab's. Whether Periodic publishes a paper rather than three blog posts; the release includes no preprint, no author list and no dataset size. Whether the company reports what Neon does to discovery throughput now that it is deployed — the justification is scientist hours freed, and that is measurable. And whether the 44.7% of hard cases Neon does not resolve get escalated to humans or quietly absorbed into the next round of synthesis decisions.
“Until then, a knowledgeable human is needed to analyse the data, compare to databases and literature, and determine what was actually made.”— Susan Latturner, Solid-state chemist, Florida State University