Databricks: GPT-5.5 Revolutionizes Enterprise Agents

Databricks has recently integrated GPT-5.5 into the workflows of its enterprise agents, marking a significant advancement in the processing of complex documents. This model has demonstrated exceptional performance on OfficeQA Pro, Databricks' internal benchmark, setting a new state-of-the-art for tasks related to enterprise documents.

OfficeQA Pro is designed to test the capabilities of models to handle parsing, retrieval, and reasoning grounded in complex environments. It includes challenges such as scanned PDFs, legacy files, and long-context documents. GPT-5.5 has managed to reduce errors by 46% compared to its predecessor, GPT-5.4, and is the first model to exceed 50% accuracy on this demanding benchmark.

Arnav Singhvi, a research engineer at Databricks, emphasizes that "Codex with 5.5 is now at the forefront of technology among all available agents and models." This improvement is particularly notable in parsing-focused workflows, where data extraction errors can have significant consequences.

The performance gains of GPT-5.5 are also evident in the orchestration of multi-step tasks. Singhvi notes that previous versions, such as GPT-5.4, tended to engage in unnecessary research detours, making processes inefficient. In contrast, GPT-5.5 has proven to be more reliable in retrieving relevant context and completing complex workflows without requiring additional supervision.

Databricks has made GPT-5.5 accessible via AI Unity Gateway, allowing clients to leverage this model in workflows built with AgentBricks and the API Agent Supervisor. This integration enables GPT-5.5 to orchestrate parsing, retrieval, and execution across specialized agents. Singhvi expresses his enthusiasm about the increasing use of these tools by clients for custom agent workflows, stating that "GPT-5.5 has been fantastic in terms of knowledge gain."