Periodic Neon beats GPT-6 Astra and Claude Fable 5.1 on XRD analysis at lower cost
zainhas · x · 2026-09-16
Periodic Labs released Periodic Neon, a model post-trained on its own lab data that outperforms GPT-6 Astra and Claude Fable 5.1 at lower cost on a challenging scientific analysis benchmark — and is now deployed in the lab analyzing real experiments in the search for better superconductors and magnets.
Key facts:
- The task is X-ray diffraction (XRD) analysis, where scientists spend hours navigating software, literature, and databases; Neon hit 55.3% success on FrontierXRD (134 lab samples), a 20x jump from the 2.7% starting point of its base model, open-weight Kimi K2.6, via midtraining and RL.
- Its in-house Periodic Harness beats off-the-shelf coding harnesses like Claude Code and Codex for these workloads.
- Evaluation used an LLM-judge ensemble (Opus 5 + GPT-5.6-Sol) calibrated against expert ratings; costs were estimated at $2.5/H200-hour for Neon and standard API prices with perfect caching for external models.
A demonstration that domain-specific post-training plus a custom harness can beat frontier general models on vertical scientific tasks.
More from Models
- GPT-6 Astra remakes Game of Thrones in low-poly Blender after 8h autonomous run — OWazabi · 2026-09-16
- ChatGPT co-inventor's startup launches Jev, a decision model claiming ultra-low hallucination — hardimanjames · 2026-09-16
- GPT-6-Astra gets "depressed" in Minecraft after creeper blows up its base, farms potatoes for hours — scaling01 · 2026-09-16
- Stripe's model-run shop bench: 5 of 7 working stores built by Claude — bcherny · 2026-09-16
- Gemini 3.8 Live undercuts speech-to-speech rivals at $0.84/hour, roughly half Gemini 3.1 Flash Live — ArtificialAnlys · 2026-09-16
- Gemini 3.8 Live hits 1.18s Time to First Audio, 2.5x faster than its predecessor — ArtificialAnlys · 2026-09-16