Databricks brings GPT-5.5 to enterprise AI agents after record OfficeQA Pro score

On May 15, 2026, Databricks announced that it is opening GPT-5.5 for enterprise agent workflows. The move follows the model's top score on OfficeQA Pro, the company's benchmark for heavy document work. On this benchmark, polished answers count for little; accuracy at every step is what matters.

Why the test matters

OfficeQA Pro doesn't measure a model's general knowledge. Instead, it tests the entire working pipeline: can the model parse a document, extract the right numbers, find the relevant passage, connect several sources, and give an answer grounded in the data? This is a sore spot for enterprise AI agents. Production systems usually break not because the model "can't think," but because it gets confused by tables, loses a number in a scan, or misreads an old PDF.

Databricks' technical report describes OfficeQA Pro as a set of 133 questions over a corpus of U.S. Treasury bulletins spanning nearly 100 years, from 1939 to 2025. The corpus contains about 89 thousand pages and more than 26 million numerical values. The set closely mimics a real enterprise environment: archives, long documents, unevenly digitized tables, outdated formats, and data where a single wrong number changes the agent's entire conclusion.

GPT-5.5's results

According to OpenAI's Databricks case study, GPT-5.5 cut errors in the agentic test harness by 46% compared with GPT-5.4 and became the first model to exceed 50% accuracy on OfficeQA Pro. In a separate post on the GPT-5.5 release, OpenAI gives a more precise figure: 54.1% on the benchmark. That's a notable jump from earlier results. In the March OfficeQA Pro report, frontier agents with direct access to the corpus averaged just 34.1%.

Databricks specifically points out that the biggest gains came in parsing-heavy scenarios. GPT-5.5 reads old documents and scanned PDFs better, extracts numbers more accurately, and less often wanders into unnecessary search detours during multistep tasks. According to the team, the model has also become more reliable both at context retrieval and at orchestrating multiple steps without additional supervision.

"With

Codex and 5.5, we got the best result among all agents and models," said Databricks research engineer Arnav Singhvi, summing up the outcome.

How it's being deployed

Databricks is now opening GPT-5.5 for customer workflows through Unity AI Gateway. The model can be used inside workflows built on Agent Bricks and the Supervisor API. According to Databricks' documentation, the Supervisor API takes some of the low-level orchestration off teams' plates: a developer specifies the model, tools, and instructions in a single request, and the platform itself runs the agent loop, calls tools, picks the next steps, and assembles the final answer.

In practice, this means Databricks isn't embedding GPT-5.5 as a standalone chat widget. It serves as a control layer on top of enterprise data and specialized subagents. Around the model, Databricks is building a typical enterprise stack:

• a single connection point for models and agents via Unity AI Gateway
• observability, rate limits, fallback routes, and request auditing
• integration with Agent Bricks, MCP servers, Unity Catalog functions, and other tools
• access controls, so users see only the sources and subagents they're permitted to use

Databricks' documentation still labels some of these components as beta, including Unity AI Gateway and the Supervisor API. But the direction is already clear: the model isn't evaluated on its own, but as a component of a managed, auditable, and secure enterprise system.

What it means

Databricks is charting a fairly pragmatic course for enterprise AI. The winner isn't simply the most talkative model, but the one that reliably reads messy documents, doesn't lose numbers, and carries a long workflow through without unnecessary errors. If GPT-5.5 holds this level in production, companies will adopt it not for demos, but to automate real document and analytics processes.