Performance on OfficeQA Pro
Databricks has announced that GPT-5.5 has set a new record on OfficeQA Pro, the company’s benchmark for complex enterprise document tasks, becoming the first model to exceed 50% accuracy on this test. In an agent-harness setting, GPT-5.5 reduced errors by 46% compared to GPT-5.4. OfficeQA Pro evaluates parsing, retrieval, and grounded reasoning across workflows that include scanned PDFs, legacy files, and long-context documents. The most significant improvements for GPT-5.5 were observed in parsing-heavy scenarios, where minor extraction errors can propagate through multi-step processes.
Integration into Agent Workflows
Databricks is preparing to make GPT-5.5 accessible to customers through the AI Unity Gateway. Customers will be able to utilize the model within workflows built with AgentBricks and the Agent Supervisor API, where GPT-5.5 will manage parsing, retrieval, and execution across specialized agents. Research Engineer Arnav Singhvi noted that earlier models, such as GPT-5.4, sometimes struggled with accurate digit extraction and would take unnecessary search detours. He attributed the enhancements in GPT-5.5 to improved parsing of older documents and more reliable context retrieval.
Expected Customer Adoption
Singhvi described the update as a measurable lift in knowledge work for organizations, stating that GPT-5.5 reduces inefficient trajectories in multi-step tasks by being more selective in searches and more consistent in completing complex workflows. Databricks positions GPT-5.5 as the preferred model for its agent systems and anticipates significant customer adoption for custom agent workflows built on AgentBricks and the Agent Supervisor API.
Original source: OpenAI News