OpenAI releases updated Astra for Law model configuration with expanded search index
Article Summary
OpenAI has released an updated Astra for Law model configuration of GPT-6.2 Astra, achieving state-of-the-art results on the Harvey LAB-AA benchmark by earning full credit on 20.8% of assignments. The updated model, which now features a legal search index expanded by over a third, increased the share of research answers passing every check on Vals AI’s Legal Research Bench from 46% to 53.8%. ChatGPT Enterprise customers, including those using the API, can now request access to the model through the Trusted Access Program.
AI Analysis
## Core Assessment
OpenAI’s updated Astra for Law configuration—built on GPT-6.2 Astra with a specialized legal search index expanded by over a third—leads the Harvey LAB-AA benchmark by completing 20.8% of tasks with full credit. On Vals AI’s Legal Research Bench, its pass rate across all checks rose from roughly 46% to 53.8%. It achieves these scores with an average of 64,000 output tokens per assignment (compared to 139,000 for Muse Wonder Full at 13.1%) and records 0.03 material hallucinations per task, matching GPT-6 and outperforming Claude Opus 5.5's 0.59 rate.
The engineering signal is clear: constraining generation length and augmenting domain-specific retrieval indexes meaningfully curbs ungrounded claims in structured legal tasks. However, the operational constraint remains severe. An industry-leading score where roughly 79% of complex assignments fail to earn full credit, alongside a 53.8% all-pass rate on research queries, proves that autonomous legal execution remains an illusion. While token efficiency saves API costs, the remaining failure rate guarantees that human legal review cannot be reduced; it simply shifts from drafting to intense verification.
## Impact and Advice
### Legal Operations and Enterprise AI Architects
Treat Astra for Law strictly as a research accelerator and drafting assistant rather than an end-to-end task runner. When requesting access via the Trusted Access Program for ChatGPT Enterprise or API workflows, mandate strict post-generation verification protocols. Do not budget for reduced associate review hours based on the 20.8% full-credit benchmark, but use the lower 64,000 average token profile to plan inference cost caps against heavier competing models.
## Bottom Line
Cutting hallucinations to 0.03 per task is impressive engineering, but when four out of five complex legal assignments still miss full marks, you are buying an eager junior paralegal who still requires adult supervision on every page.
OpenAI’s updated Astra for Law configuration—built on GPT-6.2 Astra with a specialized legal search index expanded by over a third—leads the Harvey LAB-AA benchmark by completing 20.8% of tasks with full credit. On Vals AI’s Legal Research Bench, its pass rate across all checks rose from roughly 46% to 53.8%. It achieves these scores with an average of 64,000 output tokens per assignment (compared to 139,000 for Muse Wonder Full at 13.1%) and records 0.03 material hallucinations per task, matching GPT-6 and outperforming Claude Opus 5.5's 0.59 rate.
The engineering signal is clear: constraining generation length and augmenting domain-specific retrieval indexes meaningfully curbs ungrounded claims in structured legal tasks. However, the operational constraint remains severe. An industry-leading score where roughly 79% of complex assignments fail to earn full credit, alongside a 53.8% all-pass rate on research queries, proves that autonomous legal execution remains an illusion. While token efficiency saves API costs, the remaining failure rate guarantees that human legal review cannot be reduced; it simply shifts from drafting to intense verification.
## Impact and Advice
### Legal Operations and Enterprise AI Architects
Treat Astra for Law strictly as a research accelerator and drafting assistant rather than an end-to-end task runner. When requesting access via the Trusted Access Program for ChatGPT Enterprise or API workflows, mandate strict post-generation verification protocols. Do not budget for reduced associate review hours based on the 20.8% full-credit benchmark, but use the lower 64,000 average token profile to plan inference cost caps against heavier competing models.
## Bottom Line
Cutting hallucinations to 0.03 per task is impressive engineering, but when four out of five complex legal assignments still miss full marks, you are buying an eager junior paralegal who still requires adult supervision on every page.