GPT-5.5 from OpenAI is now available in 8rney as a selectable model. Users can switch to it from Settings → Model Preferences.
What changed
GPT-5.5 is a meaningful step forward from GPT-5 on the tasks that define legal AI work. Instruction following is noticeably more reliable — the model adheres to output format specifications, word count constraints, and structural requirements more consistently, which matters when you are generating a clause in a specific form or producing a memo section that has to fit a template.
Multi-document analysis is the other area where the improvement is most evident. When a research task spans several source documents — statutes, cases, secondary material — GPT-5.5 synthesises across them more cleanly, producing summaries and comparisons that reflect the full source set rather than anchoring on whichever document appeared most prominently in context.
We also see improvement in handling ambiguous scope. Earlier GPT-5 generations would sometimes over-expand a research query, returning analysis on topics adjacent to what was asked. GPT-5.5 is better calibrated — it completes the task as specified, and where scope is genuinely ambiguous, it asks for clarification rather than assuming.
Where to use it
GPT-5.5 is well suited to tasks where output structure matters as much as output content: templated drafting, document comparison tables, structured memos, and extraction tasks where the result needs to conform to a specific schema. For open-ended research and long-form analysis, Claude Sonnet 4.6 — which remains the 8rney default — continues to perform strongly and tends to produce more discursive reasoning that legal professionals find easier to review.
Neither model dominates across all task types, which is why we offer both. Users doing varied work are best served by understanding which model to reach for when. Over the coming weeks we will publish more detailed guidance on model selection by task type.
Our approach to multi-model support
8rney supports multiple frontier models because no single model is best at everything, and because the legal AI landscape is moving fast enough that model monoculture would be a mistake. We evaluate each major release against our internal legal task suite before making it available — a set of research, drafting, extraction, and analysis tasks that reflects the actual work our users do.
GPT-5.5 passed that evaluation. It is available today for all users. As with every model we deploy, we continue monitoring quality signals in production and will update our recommendations as the picture becomes clearer.