Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-09-18 · high
Astra for Law's edge over standard Astra comes entirely from a case-law corpus and 26 legal-tool plugins, not a better model underneath.
Why now: OpenAI launched Astra for Law on September 17, giving the first benchmark numbers on what a domain corpus buys over a general model.
OpenAI released Astra for Law on September 17. It's GPT-6 Astra wired to 230 million case-law URLs covering 99.9% of published US precedent, plus 26 plugins connecting to tools legal teams already run: Relativity, Clio, iManage, Thomson Reuters.
The corpus is what moved the score. On the Vals AI Legal Research Bench, a 200-question validation set, Astra for Law hit 54.0% correctness. Base Astra with web search scored 38.7% on the same questions, same underlying model. The legal version also pulled up to 54% more relevant passages per query.
Access sits behind OpenAI's Trusted Access program. Sullivan & Cromwell, Cooley, and Latham & Watkins are already using it. API access is promised under the name gpt-6-astra-law, though OpenAI hasn't said when it opens or what it costs.
Fifteen points came from retrieval and plugins layered onto a model OpenAI already shipped. No new training run, no new architecture. That's a cheaper lever than most teams reach for first, and it's one any company sitting on a large, organized domain corpus already has access to. The gap OpenAI didn't close: whether that 54% ceiling holds once the corpus gets messier than a case-law database with a near-complete index.
Each link below shares sources, entities, or timing with this story.
Huang's September 6 post reads "From ChatGPT to o1 to Astra in 4 years. AGI has arrived," noting Astra was trained on more than 100,000 Grace Blackwell NVLink72 systems, and Greg Brockman amplified it saying OpenAI is "now moving into the AGI era" (Business Insider). Marcus re...
He argues the gains most likely come from training recipe and data, noting OpenAI's chief scientist saying compute depth stays within a factor of two of GPT-4. He walks through weight-reused looped stacks back to Universal Transformers (2018), with Nanbeige4.2-3B applying 22 b...
Ara Kharazian, who runs the Ramp AI Index, posted September 16 that OpenAI's growth is coming from shifts off GPT-5.6 Sol, off some Anthropic models, and from net-new usage, reading it as Anthropic's pace-the-frontier position costing it frontier adoption. That cuts against th...
The Astra coverage went to price and context window. The number that changes how I'd deploy it went into a system card nobody read. Artificial Analysis measured GPT-6 Astra's hallucination rate on AA-Omniscience at 51% at max effort, against 92% for its predecessor. Accuracy w...
67 on coding against Fable 5.1's 70 in Claude Code. Astra does post a 2% hallucination rate against 9.4% for GPT-5.6 Sol, and 0% scope violations against 48%. Per-task cost runs the other way, $4.72 for Astra against $9.18 for Fable 5.1 at identical $10/$50 list pricing, and A...
OpenAI's system card classifies GPT-5.3-Codex as "High" for cybersecurity — meaning it can automate end-to-end cyber operations against hardened targets. This is explicitly dual-use: the same capability that makes it an excellent security auditor also makes it an unprecedented...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.