Gemini 4 Hydrogen’s API details matter as much as its speed claims

Google introduced Gemini 4 Hydrogen as the faster, more efficient member of its Gemini 4 lineup for coding and agentic workflows. The model card says Hydrogen is adapted from Gemini 4 Carbon, accepts text, images, audio, and video, and has a combined input-output context window of up to 1 million tokens. Google also lists four thinking levels: low, medium, high, and xhigh. That makes the model’s intended role clear, but the headline context size alone doesn’t tell a developer how quickly or cheaply a particular task will run. Google says higher thinking levels can use more tokens, and its API documentation spells out implementation limits that affect real workloads.

Google’s announcement frames Hydrogen as an option for fast execution, including as a worker or verifier model alongside Carbon. The company reports a 2.5-times speed improvement over Gemini 3.8 Flash at the high setting and a Terminal-Bench 4.1 score of 57%, compared with 18.2% for Flash. Those are provider-reported comparisons. Ars Technica’s coverage also describes Hydrogen as the successor to the Flash tier and reports that the API prices are unchanged from Flash 3.8: $1.50 per million input tokens and $7.50 per million output tokens. The comparison is useful for teams already using Flash, but neither the speed figure nor a benchmark score predicts latency or cost for every prompt, tool chain, and serving setup.

The published benchmark table shows a mixed profile rather than a uniform lead. Google reports 58.8% on HLE-Diamond, ahead of the listed Claude Sonnet 5.5 and GPT-6.1 Terra scores, and 53.4% on its Multimodal Knowledge Agent benchmark. On Terminal-Bench 4.1, however, Hydrogen’s 57% trails Sonnet 5.5’s 63.9%; on OSWorld-2.1, its 77.6% is below the listed 83.9% for Sonnet. These are results in Google’s model card, and its announcement describes the relevant evaluations as company-led. The scores help identify where Hydrogen may be worth testing, but they don’t establish a general ranking across different tasks or independent reproductions.

For API users, the fine print changes how Hydrogen can fit into an application. Google Cloud documents support for implicit context caching only, with a 4,096-token minimum; explicit caching is not supported. Custom temperature, top-p, and top-k values are ignored, and requesting a minimal thinking level or setting a thinking budget returns an HTTP 400 error. The service can extend output to the full context limit through long decoding, but that uses continuation handling or background requests, so it isn’t the same as a single ordinary synchronous response. The model card also notes possible hallucinations, occasional slowness or timeouts, and a March 2026 knowledge cutoff. Those limits make a small workload-specific evaluation more useful than choosing on context size or vendor benchmark claims alone.