GPT-5.5: Research Preview Results
A first look at GPT-5.5 from Harvey.
Today, we're sharing early access results for GPT-5.
GPT-5.5 is OpenAI’s latest frontier model, building on the strengths of GPT-5.4 with improved substantive accuracy, stronger organizational structure, and more consistent formatting across legal practice areas. This model is currently in research preview.
Early access evaluations show GPT-5.5 delivering gains across both transactional and litigation tasks, with particular strength in risk assessment, deal management, and analysis of litigation filings. On our BigLaw Bench evaluation suite, GPT-5.5 scored 91.7%, up from GPT-5.4’s 91.0%. This is one of the highest scores we’ve seen to date. The model achieved 43% perfect scores and 87% of tasks scored above 0.80, with zero scores below 0.50.
Our teams highlighted several additional strengths. GPT-5.5 showed notably improved output organization and readability, with evaluators praising its use of structured layouts. On drafting tasks, evaluators noted that GPT-5.5 delivered outputs in proper format with intuitive structure, and the model made more effective use of bold headings and citations grounded in source documents.
“GPT-5.5 shows improvement in legal reasoning, organizational structure, and audience calibration. These are the kind of practical gains that move the needle for legal practitioners. This performance also translates to benchmark scores, with GPT-5.5 scoring 91.7% overall with perfect scores on 43% of tasks.””
Niko Grupen
Head of Applied Research at Harvey
As we’ve seen with prior models in the GPT family, verbosity remains an occasional challenge — some responses were more detailed than the task called for, particularly on straightforward queries. GPT-5.5 is currently in research preview. We'll be rolling it out to eligible clients once it becomes generally available.
