1300张卡叫板10万卡！前OpenAI副总裁出手，硬核难题击败GPT-6 Astra

Periodic Labs' ~1T-parameter model Periodic Neon, built by ex-OpenAI VP Liam Fedus, hit 55.3% on FrontierXRD (134 hard X-ray diffraction analysis problems), beating GPT-6 Astra (~7x params, 100K+ Grace Blackwell GPUs) and Claude Fable 5.1 (~40%) at ~$4/problem vs ~$7 for Astra, using just 1300 H200 GPUs. The edge comes from RL training on proprietary XRD data generated by their 24/7 Menlo Park materials lab - replacing exhausted internet text with a self-renewing "data well" - plus a custom Periodic Harness tool environment (3.8x success rate vs Claude Code-based baselines) and LLM judges (Opus 5 + GPT-5.6 sol) calibrated to 84% agreement with expert consensus. Efficiency gains stem from decoupled training/inference GPU isolation, a pbox sandbox with 3.3x throughput, and 95% cluster utilization.

why now

前OpenAI副总裁新模型用1300卡击败GPT-6 Astra，数据飞轮新路径。

topic

AI Tech & New Models

source

新智元