AI HOT (Curated Pool)May 15, 2026, 08:00 UTC+8

Databricks brings GPT-5.5 to enterprise agent workflows

Databricks将GPT-5.5引入企业智能体工作流

Databricks made GPT-5.5 available through AI Unity Gateway for AgentBricks and Agent Supervisor API workflows; on OfficeQA Pro, it became the first model above 50% accuracy and reduced errors by 46% versus GPT-5.4.

Why it matters: HKR-H/K/R all pass: GPT-5.5 enters Databricks workflows with 50% OfficeQA Pro accuracy and 46% fewer errors than GPT-5.4. It stays below a full model-release score because the page is a sales-led OpenAI customer story using Databricks’ own benchmark.

Read the original ↗Export Markdown

Take

GPT-5.5’s useful gain here is not abstract reasoning; it is better behavior on ugly enterprise inputs. Databricks says the model crossed 50% accuracy on OfficeQA Pro and cut errors by 46% versus GPT-5.4. The named failure mode matters: scanned PDFs, legacy files, long-context documents, digit extraction, and fewer pointless retrieval detours.

I buy half the story. OfficeQA Pro is Databricks’ own benchmark, and the article does not disclose sample size, task mix, or the full model leaderboard. Also, 50% accuracy is still a messy production bar. But the distribution path is real: AI Unity Gateway, AgentBricks, and Agent Supervisor API put GPT-5.5 directly inside enterprise agent workflows. Most RAG agents have looked fine in demos and then bled out on document grime. If 5.5 really reduces extraction cascades, that is a more monetizable improvement than another chat leaderboard bump.

Source: AI HOT (Curated Pool) · openai.com
Published: May 15, 2026, 08:00 UTC+8 · May 15, 2026