GPT-5.5 is SOTA for Databricks

The benchmark in question, OfficeQA, is an internal or domain-specific evaluation tied to enterprise document workflows, not a widely audited third-party suite. A 46% error reduction measured on a task distribution that OpenAI and Databricks jointly benefit from publicizing deserves more scrutiny than the announcement invites.

The timing here is notable given Hugging Face's recent piece on how 'AI evals are becoming the new compute bottleneck.' That story argued evaluation infrastructure is now a credibility signal, not just a development tool. This announcement does the opposite of what that piece recommends: it leans on a narrow, commercially adjacent benchmark rather than broad, independently administered evals. The result is a capability claim that is difficult to verify or compare against anything outside the Databricks context. That gap matters more as enterprise buyers grow sophisticated about distinguishing marketing benchmarks from production performance.

Watch whether Arnav Singhvi or Databricks publish OfficeQA methodology and test-set composition publicly within the next 60 days. If they do not, the 46% figure has no external reference point and should be treated as a product marketing number rather than a reproducible result.