Databricks brings GPT-5.5 to workflows

OpenAI and Databricks said on May 15 that GPT-5.5 is being used in Databricks enterprise agent workflows, extending a partnership that already put OpenAI models inside the company’s data and AI platform. The announcement was framed around sales, operations and coding tasks that run across multiple tools rather than around a standalone chatbot. Databricks said all GPT-5.5 usage on its platform is governed through Unity AI Gateway, which it described as a control layer for security, cost management and observability. OpenAI said Databricks adopted the model after GPT-5.5 posted a new high score on OfficeQA Pro, a benchmark OpenAI cited for workplace question answering. ### Which product actually changed on Databricks? Databricks documentation updated in April said OpenAI GPT-5.5 and GPT-5.5 Pro are available as Databricks-hosted models through Mosaic AI Model Serving. The supported-models page lists the endpoint name as `databricks-gpt-5-5` and says the model accepts text and image inputs. The same page says GPT-5.5 has a 400,000-token total context window and a maximum output of 128,000 tokens. (databricks.com) A Databricks blog post published on May 18 said customers can use GPT-5.5 directly on Databricks for coding workflows with Codex, enterprise agents and document pipelines. The company said governance runs through Unity AI Gateway from the start, covering “agents, queries, and coding workflows.” ### Why are both companies talking about “agent workflows” instead of chat? (docs.databricks.com) OpenAI said on April 23 that GPT-5.5 was built for “complex tasks like coding, research, and data analysis across tools,” and described the model as able to plan, use tools, check its work and keep going on multi-part tasks. The API documentation tells developers to use the Responses API for reasoning, tool-calling and multi-turn work, and says higher reasoning settings are intended for “complex agentic tasks.” (databricks.com) Databricks used similar language in its own materials. A Databricks blog post from April 24 called GPT-5.5 OpenAI’s strongest frontier model for “agentic work in enterprise,” complex document reasoning and long-horizon coding agents. The company’s October partnership announcement said enterprises could call OpenAI models from SQL, model-serving endpoints or Agent Bricks to build domain-specific agents on enterprise data. (openai.com) ### What does Databricks say sales teams can do with it? OpenAI’s Databricks case-study page said sales teams can use Codex to generate pipeline briefs, meeting-preparation packets, forecast reviews, account plans and diagnoses for stalled deals from work inputs already inside business systems. The page quotes Databricks research engineer Arnav Singhvi saying, “Having GPT-5.5 supervise these workflows is really exciting,” and says the model delivered a “knowledge lift” for knowledge-work tasks. (databricks.com) The Databricks side of the rollout ties those use cases to orchestration features that were added in April. Release notes published last week said the Databricks SDK for Python can create and manage Supervisor Agent tools, and that Supervisor Agent now supports custom MCP servers and custom agents hosted in Databricks Apps. Databricks’ agent authoring documentation, updated May 12, says developers can build and deploy custom agents through Databricks Apps with control over code, server configuration and deployment workflow. (openai.com) ### How is Databricks trying to make the model acceptable for enterprise use? Databricks said governance is the selling point alongside model capability. The company said Unity AI Gateway provides centralized security controls, cost controls and observability for GPT-5.5 usage on the platform. Earlier Databricks materials said the same gateway can apply rate limits, monitor quality and enforce guardrails such as PII detection across model providers. (docs.databricks.com) OpenAI’s API pricing page adds a cost detail that matters for large workflow runs: GPT-5.5 is priced at $5 per 1 million input tokens and $30 per 1 million output tokens, with higher charges for very large prompts and a 10% uplift for regional processing endpoints. Those figures sit alongside Databricks’ own controls for routing and usage management. ### What should readers watch next? (databricks.com) April 30 is the date Databricks gave for Supervisor Agent support for custom MCP servers and custom agents on Databricks Apps, and May 12 is the latest documentation update for authoring and deploying custom agents. The next concrete milestone is broader customer use of those tools with GPT-5.5 and GPT-5.5 Pro inside Model Serving and Agent Bricks, according to Databricks product pages and release notes. (docs.databricks.com) (developers.openai.com)