Google says Gemini in Sheets reaches state-of-the-art 70.48% on SpreadsheetBench

What Google highlighted on X

On March 10, 2026, Google AI said Gemini in Google Sheets reached 70.48% on the full SpreadsheetBench dataset. Google characterized that result as state of the art and said it comes close to human expert performance, which is a more ambitious claim than the usual spreadsheet-assistant language around formulas or autofill.

The company also used the X thread to place Sheets inside a broader Workspace strategy. In the lead post, Google said Gemini had to be tailored to each product's data model and user needs across Docs, Sheets, Slides, and Drive. The Sheets-specific post suggests Google is pushing beyond generic chat inside productivity apps toward domain-tuned behavior for structured business work.

What the Google blog adds

Google's supporting blog post says the 70.48% score was achieved by equipping Gemini in Sheets with stronger verbalization and coding capabilities. According to Google, that allows the model to natively build complex models and dashboards, solve optimization problems, and verify its own work rather than only generate spreadsheet formulas on request.

• Google says the result exceeds competing systems on the full SpreadsheetBench benchmark.
• The company also says the score approaches human expert ability on spreadsheet tasks.
• Google frames the improvement as part of product-specific model adaptation inside Workspace, not just a general Gemini model upgrade.

Why this matters for enterprise workflows

Spreadsheet tooling is one of the clearest enterprise tests for whether AI can move from assistant behavior to delegated analytical work. Formula suggestions are useful, but real business value appears when a system can understand a table, reason through a goal, construct a model, and check whether the output is actually coherent. That is the transition Google is signaling here.

If the benchmark gains hold up in production, the implication is broader than Sheets alone. Product-specific tuning inside Workspace suggests Google sees AI productivity gains as a data-model problem as much as a language-model problem. In practice, that means future competition may revolve less around generic model IQ and more around how deeply vendors integrate with the native structures of documents, spreadsheets, presentations, and knowledge stores.

Sources: Google AI X post, Google Workspace blog