Krux
May 18, 2026
GPT-5.5 Cuts Enterprise Task Errors in Half
Published: May 18, 2026 at 12:11 AM
Updated: May 18, 2026 at 12:11 AM
100-word summary
Databricks is now offering GPT-5.5 through OpenAI for businesses building automated agents. The model slashed errors by 46% versus its predecessor on OfficeQA Pro, a benchmark testing whether AI can parse scanned documents and execute multi-step tasks without wandering off script. It's the first to crack 50% accuracy on that test. The upgrade matters most for document-heavy workflows where small parsing mistakes compound fast. Companies can access it through Databricks' AI Unity Gateway for workflows stringing together multiple specialized agents. OpenAI cautions the benchmark ran in a research setting, so real-world mileage may vary.
What happened
Databricks is now offering GPT-5.5 through OpenAI for businesses building automated agents. The model slashed errors by 46% versus its predecessor on OfficeQA Pro, a benchmark testing whether AI can parse scanned documents and execute multi-step tasks without wandering off script. It's the first to crack 50% accuracy on that test. The upgrade matters most for document-heavy workflows where small parsing mistakes compound fast. Companies can access it through Databricks' AI Unity Gateway for workflows stringing together multiple specialized agents.
Why it matters
OpenAI cautions the benchmark ran in a research setting, so real-world mileage may vary.