Errors down 46%: GPT-5.5 lands in Databricks agent workflows!

Hey everyone, I'm Shii-chan! Today I've got a fun one about AI agents working in the enterprise trenches.

OpenAI News openai.com

What was announced?

This comes from OpenAI's News. Databricks — the data and analytics platform — is bringing GPT-5.5 to its customer agent workflows!

The trigger was GPT-5.5 setting a new state of the art on OfficeQA Pro, Databricks' own benchmark. OfficeQA Pro measures how well a model can parse, retrieve, and do grounded reasoning over scanned PDFs, legacy files, and long-context documents — exactly the kind of tasks that tend to break production agent systems.

Databricks research engineer Arnav Singhvi puts it like this:

"Codex with 5.5 is now state-of-the-art amongst all the agents and models out there."

The story so far

With the previous GPT-5.4, there were rough spots. On older documents and scanned PDFs, it sometimes couldn't read the digits correctly.

As Singhvi explains, once you can't extract a certain digit, the whole trajectory of what the agent works with changes — one small parsing miss cascades downstream through the rest of the workflow. GPT-5.4 also sometimes took unnecessary search detours on multi-step tasks, leading to inefficient trajectories.

What changes

GPT-5.5 improves this parsing story a lot. In the agent-harness setting, GPT-5.5 reduced errors by 46% compared to GPT-5.4, and it became the first model to surpass 50% accuracy on OfficeQA Pro.

Databricks saw the biggest gains in parsing-heavy workflows, describing a step-function lift on older documents and scanned PDFs. It was also more reliable at retrieving relevant context and finishing complex workflows without extra supervision.

Dive Deep

So how does it slot into production? Databricks now offers GPT-5.5 through AI Unity Gateway, and customers use the model inside workflows built with AgentBricks and the Agent Supervisor API. Here GPT-5.5 acts like a conductor, orchestrating parsing, retrieval, and execution across specialized agents.

"GPT-5.5 has been great in terms of knowledge lift. It's a step size function change in terms of doing knowledge work for us."

Having GPT-5.5 supervise these workflows looks like the shape of custom agents to come.

Wrap-up

• Databricks brings GPT-5.5 to customer agent workflows (OpenAI's News)
• On its OfficeQA Pro benchmark: 46% fewer errors vs GPT-5.4, and first to pass 50% accuracy
• Biggest improvements in parsing older documents and scanned PDFs
• Delivered via AI Unity Gateway, with GPT-5.5 as the conductor in AgentBricks and Agent Supervisor API workflows

If you're building enterprise agents to automate paperwork-heavy work — or you're just curious how much GPT-5.5 actually helps in the field — this one's for you!