Back Home

最新模型

Google Releases Gemini 4 Hydrogen with Four Reasoning Levels and Dynamic Agent Model Switching

Google released Gemini 4 Hydrogen on November 12 and added an Orchestrator to Antigravity for dynamically switching models. The new version offers a unified context window of up to one million tokens, but independent evaluations at the highest reasoning level show that a lower unit price does not necessarily mean shorter task times.

Gciriani · CC BY-SA 4.0 · Image source
zh-Hant

Google launched Gemini 4 Hydrogen on November 12 for everyday coding, tool use, and multimodal agent tasks, making it available through channels including the Gemini API and AI Studio. Input and output are priced at $1.50 and $7.50 per million tokens, respectively, at the same rates across the entire context window. Reasoning settings are low, medium, high (the default), and the newly added xhigh. Official announcement

Developers should first clarify the output limits. The model card says Hydrogen is adapted from Carbon and uses a shared input and output window of up to one million tokens across text, images, audio, and video. Synchronous output is capped at 64K tokens by default; generating up to one million output tokens requires background long decoding. In practice, long document inputs and long answers compete for the same budget, so a million-token context window should not be interpreted as allowing a synchronous answer of the same length. Model card

The Antigravity Orchestrator, launched at the same time, can dynamically switch between Carbon and Hydrogen within a single work session. Google says that on Terminal-Bench 4.0, this configuration came close to Carbon’s task success rate while using 24% less quota on average and reducing completion time by 19%. This provides a way to assign planning and execution to different models, but the figures come from a specific Google test, and the quota reduction should not be treated as an equivalent reduction in API bills. Official details

Independent evaluations show the cost of higher reasoning intensity. The xhigh version scored 54 on the Intelligence Index from Artificial Analysis, but took an estimated 13.5 minutes per question, slightly above the 13-minute median for the comparison group in the same price tier. It also generated 304 million tokens across the full evaluation, more than the comparison group’s median. These results apply only to this setting and evaluation workload; they should not be generalized to all online requests. Independent evaluation

Teams considering adoption should next compare success rates, retry counts, total tokens, and end-to-end latency across reasoning settings using tasks from their own codebases, then decide on a routing strategy. The model card still lists hallucinations, timeouts, and increased token usage at higher reasoning settings among its limitations, and a full safety report has yet to be released. Teams should retain tool permission controls and result verification when selecting a model. Limitations

Sources

  1. Introducing Gemini 4 Hydrogen
  2. Gemini 4 Hydrogen — Model Card
  3. Gemini 4 Hydrogen (xhigh) — Intelligence, Performance & Price Analysis