Three provider moves land today. OpenAI locks Astra for law firms. Anthropic opens verified biology access. NVIDIA posts a Vera Rubin preview on MLPerf.
In the news
OpenAI configures Astra for law firms
OpenAI published Astra for Law on September 17. The page describes GPT-6 Astra with a legal search index and instructions for legal analysis and writing. The index covers U.S. case law, statutes, regulations, court rules and administrative decisions. It claims more than 230 million URLs with daily updates and more than 99.9 percent coverage of published U.S. precedential case law via Free Law Project and CourtListener.1
OpenAI claims a private Vals AI Legal Research Bench test with 200 U.S. questions. Astra for Law scored 54.0 percent overall correctness versus 38.7 percent for GPT-6 Astra with web search. The company says that is a 40 percent relative gain. Case law questions show 24 percent more reference cases and up to 54 percent more relevant passages from correct opinions.
Access is via Trusted Access in ChatGPT and Codex for selected firms. The API version is coming soon as gpt-6-astra-law. Named partners on the page include Harvey and Legora. The page quotes John Savva of Sullivan & Cromwell and notes governance work with Latham & Watkins.
Anthropic opens biology access with verification
Anthropic announced the Life Sciences Verification Program on September 17. The program gives verified life science teams Mythos, Opus and Sonnet with more permissive biology classifiers than generally available Fable models. It is in beta for teams and institutions. Individual Pro and Max access is planned later.2
Standard Use covers most life science work, team wide, renewed annually. It applies to Mythos 5.1, Opus 5 and Sonnet 5 today and to future models. High risk Use is an add on for one research project, renewed every six months, and removes remaining life science blockers.
High risk is available for Opus 5 and Sonnet 5 today. Mythos high risk is limited and requires extra vetting and U.S. government coordination. Cyber classifiers stay on.
Monitoring shifts from real time blocking to offline monitoring against the grant stated use case. Data retention for LSVP traffic is 30 days. Data is not used for training and is not accessible to Anthropic life science researchers. Early access names include Xaira, Edison and Manifold Bio.
Vera Rubin NVL72 appears in MLPerf preview
NVIDIA posted a Vera Rubin NVL72 MLPerf Inference v6.1 preview on September 16. Results were retrieved from mlcommons.org on September 16. Entries 6.1-0106 and 6.1-0074 show a preview submission for Vera Rubin NVL72.3
NVIDIA claims up to 3.7x throughput on Qwen3-VL versus GB300 NVL72 with vLLM and NVIDIA Dynamo in offline, server and interactive modes. It claims up to 2.5x on DeepSeek-R1 with TensorRT-LLM. The submission uses NVFP4, disaggregated prefill decode and expert parallelism. Nebius also submitted a Vera Rubin preview.
GB300 scaling is also shown. Four racks with 288 GPUs reach 99 percent scaling efficiency offline versus one 72 GPU rack. Entries 6.1-0073 and 6.1-0074 are cited. WAN 2.2 text to video is claimed at 0.65 720p videos per second at 5.7 seconds per video, 9x versus a single node.
