What changed
Google released Gemini 4 Hydrogen globally across the Gemini API, AI Studio, Gemini for business, the Gemini app, Search, Sheets and paid Antigravity plans. The model combines a one-million-token unified input-and-output window with background generation and a continuation token for long tasks.
Why it matters
A lower-cost Gemini 4 model with long background generation creates a concrete option for coding, document and agent workloads that outgrow ordinary response windows. The unified limit also means long output directly consumes the same budget as input.
What to do
Canary one long-context task against the current model, cap input plus output together, persist continuation tokens and compare latency, quality and total cost before changing production routing.
Caveat
Google's quality, coding, safety and efficiency results are provider claims. Background output and continuation semantics need workload tests, and Google's enterprise documentation allows short-term GA models to retire with 45 days' notice.