[Kamilbenkirane]

What

Add gemini-4-hydrogen (Gemini 4 Hydrogen) after gemini-4-carbon, served by the Interactions backend (API key; Vertex also lists it GA with no allowlist): MAX_TOKENS Range(min=1, max=65536), THINKING_LEVEL ["low", "medium", "high", "xhigh"], WebSearch/CodeExecution/UrlContext, ToolChoiceSupport(), Schema(), image/video/audio/document input, streaming=True.

Why

Google released gemini-4-hydrogen GA on the Gemini API, free and paid tiers, on November 12, 2026; Google Cloud lists it GA the same day.

Evidence

• gemini-4-hydrogen (ModelOption), ThinkingLevel low/medium/high/xhigh (plus minimal), max_output_tokens, google_search/code_execution/url_context, tool_choice, response_format, image/video/audio/document content: Interactions API reference
• Gemini 4 Hydrogen; text/image/video/audio/PDF in, text out; max output 65,536 synchronous, 1,048,576 only with long decode; thinking low, medium, high, xhigh; search, code execution, URL context, function calling, structured outputs: Gemini 4 Hydrogen
• Rejects minimal; output past 65,536 needs background; free and paid tiers: What's new in Gemini 4 Hydrogen; GA November 12, 2026: Release notes
• XHIGH in ThinkingLevel: GenerateContent API reference; thinking_level="MINIMAL" returns HTTP 400, long decode on v1beta1, no allowlist: Gemini 4 Hydrogen (Agent Platform); GA November 12, 2026: Model versions and lifecycle
• Conflict: the reference's generic ThinkingLevel also lists minimal; live serving rejected it (call below) and wins. streaming=True: the stream call below.

Files

• src/celeste/modalities/text/providers/google/models.py

Validation

make ci: pass
Live call: made (Celeste Interactions stream, thinking_level="low", max_tokens=16: 2 chunks, status incomplete with usage parsed and no text, the cap spent on 13 reasoning + 3 output tokens; unary thinking_level="minimal" via extra_body: rejected with "Thinking level THINKING_LEVEL_MINIMAL is not supported for this model")

Depends on

none