Google appears to be preparing an "xhigh" thinking level for the Gemini API; official SDKs add a value seemingly above high in an unannounced leak

Changes in official SDKs, no announcement from Google

Google is moving to add a new value, "xhigh," to the setting in its developer-facing Gemini API that controls how much an AI model thinks before giving an answer (reasoning depth, or thinking level). On Nov. 8 U.S. time (4:02 to 4:03 a.m. Nov. 9 Japan time), proposed changes (pull requests) adding xhigh to the reasoning depth of the "Interactions API," which calls models and agents, were filed on GitHub for the Python and JavaScript versions of the company's official SDKs (software development kits). Both were titled "feat: add xhigh thinking level to Interactions ThinkingLevel" and were posted by copybara-service, an automation account Google uses to sync code.

Google has not announced the value on its official blog or official X accounts, and the information remains a leak about an unannounced feature. ORIENS confirmed on Nov. 12 that neither Google's official blog nor posts since Nov. 1 from five official Google-related X accounts included any announcement of xhigh.

@aitrackerbot, a bot that detects API changes, reported it on X at 11:14 a.m. Nov. 9 Japan time. The post said the same changes had also been proposed for the Go SDK and the "Gemini API CLI," and that they described xhigh as "Extra high thinking level, above high." ORIENS has not confirmed the Go and CLI changes or this description in official materials. The Python type definition does not contain this description.

Reasoning depth goes to 5 levels

According to a commit to the official Python SDK (Nov. 9 U.S. time, 4:40 a.m. Nov. 10 Japan time), the change adds a single xhigh line to the type definition file for the Interactions API. In the definition at that point, there were five values that could be specified for reasoning depth: "minimal," "low," "medium," "high" and "xhigh."

The values are listed from shallowest to deepest, with xhigh placed last. Judging by the name and the order, it appears to be preparation for adding a level above high, which had been the deepest. However, the definition file has no description for each value, and Google has not officially explained what xhigh means.

In Google's official SDKs, the "Decisions API," which returns only the answer to a choice or judgment, was also found as a proposed change before it was announced.

Proposed changes shown as "open," but listed in release notes

Python proposal #3320 and JavaScript proposal #2159 were both shown as "open" when ORIENS retrieved them on the morning of Nov. 12 Japan time. Meanwhile, a change with the same title was merged into the Python version as a commit, and "add xhigh thinking level to Interactions ThinkingLevel" appeared in the release notes for Python SDK 3.1.0, JavaScript SDK 3.1.0 and Java SDK 2.1.0, dated Nov. 10.

The JavaScript and Java release notes were published on Nov. 10 U.S. time (5:17 a.m. and 5:18 a.m. Nov. 11 Japan time), after @aitrackerbot's post. What appeared in the release notes is a value that can be specified in the SDKs, and ORIENS has not confirmed whether the API side accepts xhigh.

No word on supported models or pricing

The official SDK changes and release notes do not state which models support xhigh, its pricing, when it will be offered, or its token usage or response speed. @aitrackerbot's post also said no supported model has been named.

The official Gemini API documentation explaining reasoning depth could not be loaded when ORIENS tried to retrieve it on Nov. 12, so whether it mentions xhigh has not been confirmed. ORIENS has not run tests with an API key, and it is not known whether calls specifying xhigh currently work.