OpenAI Launches GPT 6 Astra Enhancing Legal AI Capabilities
OpenAI has introduced GPT 6 Astra, its latest flagship model and the most advanced large language model to date for end to end tasks. The rollout commenced on September 3rd, initially on a limited basis, with plans for broader availability in the future.
The release of GPT 6 Astra is significant as it demonstrates the ongoing improvements in AI performance by major model developers. This progression is particularly relevant for legal professionals, as there exists a threshold below which legal AI is of limited utility, but once a certain level of reliability is achieved, it can become immensely beneficial.
OpenAI has shared various benchmarks indicating substantial enhancements, including the 'Agents Last Exam' benchmark from UC Berkeley RDI, which evaluates whether an AI can perform economically valuable professional tasks from start to finish. GPT 6 Astra achieved a score of 59.3%, compared to 53.6% for its predecessor, GPT 5.6 Sol.
While the current performance may not fully replace entire streams of legal work, the rapid improvement is noteworthy. The initial scepticism surrounding the capability of generative AI to handle complex legal tasks has evolved significantly, with increasing competition among model developers and open source contributors.
OpenAI has stated that GPT 6 Astra integrates advancements in computational power with specialised training for professional environments, enabling it to manage complex tasks and produce refined documents. The model is designed to enhance understanding of user intent and improve task delegation, which is crucial for a more comprehensive approach to legal work.
Harvey's Niko Grupen, Head of Applied Research, remarked on the significant quality enhancements of Astra over GPT 5.6 Sol, particularly in complex legal tasks. Early testing indicated that Astra approaches legal work similarly to a discerning lawyer, effectively distinguishing between documents and identifying unsupported assumptions.
Legora also conducted tests on GPT 6 Astra, focusing on a complex workflow involving financial statement tie outs. The model successfully processed 41 documents in a single run, identifying discrepancies and providing a detailed record for legal professionals to review. This capability highlights the model's efficiency and the confidence it instils in its results.
Legora's evaluation using its Benchmark for Agentic Reasoning indicated that GPT 6 Astra improved performance by nearly 40% on the financial statement workflow, with an average improvement of about 3% across all tasks. This progress suggests that advancements in AI models can occur meaningfully within short timeframes.
While GPT 6 Astra has not yet reached a level of performance that allows for complete autonomy in legal workflows, the notable rate of improvement raises questions about the future role of AI in legal practice. The ongoing enhancements could lead to a shift in how legal work is conducted, particularly regarding the balance of human oversight and AI capabilities.