LLMs

OpenAI's Astra for Law tops legal reasoning benchmark with efficiency gains

GPT-6.2 Astra configured for law achieves 20.8% full-credit pass rate on Harvey's LAB-AA, outpacing competitors while using less compute per task.

Last verified:

According to the OpenAI Blog, Astra for Law—a specialized configuration of GPT-6.2 Astra paired with an expanded legal search index—has achieved top marks on Harvey’s Legal Agent Benchmark (LAB-AA) and competing legal reasoning tests. The model earned full credit on 20.8% of assignments on LAB-AA, outperforming Muse Wonder Full, the highest-scoring non-OpenAI competitor, by nearly 40% more often. OpenAI is now rolling out access to all ChatGPT Enterprise customers through the Trusted Access Program, as well as via the API.

Benchmark Performance and Competitive Position

Astra for Law’s LAB-AA performance places it substantially ahead of competing models on a rigorous legal reasoning task. The benchmark, co-developed by Harvey and Grubel, presents models with 120 real-world legal assignments spanning 24 practice areas—from drafting deposition outlines to reviewing commercial leases to structuring acquisitions. To earn full credit, a model must meet every requirement without introducing unsupported factual claims.

According to OpenAI, Astra for Law achieved 20.8% full-credit assignments, compared to 15.0% for base GPT-6.2 Astra and 13.1% for Muse Wonder Full. The result reflects both the base model’s capabilities and the domain-specific tuning applied through legal search indexing and instruction design.

Performance on other legal reasoning benchmarks reinforces this trajectory. According to OpenAI’s report, Astra for Law is a top performer on LAB: USCL and on Vals AI’s Legal Research Bench. On the latter, OpenAI notes the model’s pass rate—the share of research answers passing every check—rose from approximately 46% to 53.8% in less than two months following an expansion of the legal search index by over one-third.

Deployment and Market Access

Astra for Law entered limited availability in September 2026 with select law firms including Cooley, Ropes & Gray, and Sullivan & Cromwell. The OpenAI Blog reports that the model has undergone rapid iterations to strengthen legal reasoning and broaden its knowledge base in that interval.

The November 2026 expansion to ChatGPT Enterprise marks a shift toward broader adoption. Access is offered through the Trusted Access Program, requiring customers to request authorization rather than automatic enablement—a practice OpenAI uses to manage demand and gather feedback before full deployment.

Why This Matters

For legal practice, model accuracy on hallucination-prone tasks is a licensing issue: a lawyer cannot ethically deploy a system that generates fabricated case law or misrepresents contract terms. Astra for Law’s strong performance on LAB-AA signals that domain-specific tuning of frontier models can reduce this risk substantially.

The timing matters as well. Law firms face mounting pressure to absorb AI-driven efficiency gains—generalist models have proven useful for drafting and research acceleration, but their error rates on factual recall make them unsuitable for unsupervised deployment. A model explicitly calibrated for legal reasoning may shift the calculus for mid-market and enterprise firms deciding whether to build internal tooling or license third-party platforms like Harvey or LexisNexis+ AI.

The expansion to ChatGPT Enterprise also signals OpenAI’s strategy to compete directly with legal-focused platforms by positioning its frontier model as the substrate, with configuration layers—rather than a separate product line—as the differentiator. This approach mirrors OpenAI’s earlier moves in domain-specific applications but centers the model family, not the vertical app.

Frequently Asked Questions

What is Astra for Law and how does it differ from base GPT-6.2 Astra?

Astra for Law is a specialized configuration of GPT-6.2 Astra that pairs the base model with an expanded legal search index and instructions tuned for legal research, reasoning, and writing. According to OpenAI, the legal-specific configuration performs significantly better on legal benchmarks than the base model alone.

What is LAB-AA and why does it matter?

LAB-AA (Artificial Analysis's version of Harvey's Legal Agent Benchmark) tests whether a model can complete full legal assignments across 24 practice areas without hallucinating facts. It includes 120 real-world assignments such as drafting deposition outlines and reviewing commercial leases, making it a rigorous measure of legal reasoning fidelity.

How much has Astra for Law improved since September?

According to OpenAI, the pass rate on Vals AI's Legal Research Bench improved from approximately 46% to 53.8% in under two months, following expansions to the legal search index and improvements to reasoning instructions.

Who can access Astra for Law?

OpenAI is now enabling all ChatGPT Enterprise customers to request access through the Trusted Access Program, including via the API. The model was initially made available to select law firms including Cooley, Ropes & Gray, and Sullivan & Cromwell in September.

#gpt-6 #legal-ai #benchmarks #enterprise