python-genai v3.3.0

Version 3.3.0 of google-genai, the Python library Google publishes for calling its Gemini models, mostly brings the library in line with Gemini 4 Hydrogen, the Gemini 4 model Google DeepMind built for fast everyday coding and agent work and announced on November 12, the same day this release is dated. According to the release notes, it adds an XHIGH value to the SDK's ThinkingLevel setting and adds gemini-4-hydrogen to the list of model names used by the SDK's Interactions types. The ThinkingLevel commit places XHIGH directly after HIGH and includes a test showing that the lowercase string "xhigh" maps to the same value.

The thinking level decides how much a Gemini model reasons internally before it answers. Google's thinking documentation calls it the main control over quality, latency and token use, and says xhigh works only on gemini-4-hydrogen, which defaults to high. Google's announcement reports that Hydrogen scores 51 on the Artificial Analysis Intelligence Index at high and 54 at xhigh. The same documentation warns that xhigh raises latency and token use and can cause overthinking or use up the max_output_tokens limit on simple prompts. Because Google bills the full thinking tokens even though the API returns only a summary of them, xhigh is a setting to choose for individual hard requests rather than a better default.

The release also adds four reasons a response can stop to the SDK's FinishReason list, including MISSING_THOUGHT_SIGNATURE and MODEL_ARMOR, the latter for a response blocked by Google's Model Armor filter. The commit marks three of them as unavailable on Vertex AI and MODEL_ARMOR as unavailable on the Gemini API, so code that branches on why generation ended can now name those cases. The README was also updated to describe the breaking changes from version 3.0, which made v1 the default Gemini Developer API version and switched the default HTTP client to httpx2.