Artificial Analysis's Intelligence Index v5.0.1 scored Gemini 4 Hydrogen (xhigh) at 54, while the same model scored 51, 41, and 27 at high, medium, and low, respectively. The index is a weighted aggregate across multiple tasks and does not represent performance on every task.
Tasks this can help assess: Overall capability across knowledge work, coding, science, and general knowledge.
Tasks this should not be generalized to: Tasks not included in the index, workflows in specific industries, and non-English or purely visual tasks.
Applicable model version: Gemini 4 Hydrogen, with the xhigh, high, medium, and low reasoning tiers reported separately.
Test environment or client: Independent evaluation by Artificial Analysis; the organization did not disclose which model endpoint it used for this evaluation.
Reasoning tier and parameters: Results are reported separately by reasoning tier. The organization's general settings are temperature 0.6 for reasoning models (unless the model provider recommends otherwise) and an output limit set to the maximum disclosed by the provider.
Intelligence Index v5.0.1 includes 9 evaluations across four categories, each contributing 25% of the total score:
| Category | Evaluations and item counts | Repetitions | Share of total score |
|---|---|---|---|
| Knowledge work | AA-Briefcase v1.1 (91 items), AutomationBench-AA (657 items), GDP.pdf (100 items) | 1, 1, 5 | 15%, 5%, 5% |
| Coding | FrontierCode-AA 2.0 (132 items), Terminal-Bench 4.0 (66 items) | 3, 3 | 15%, 10% |
| Science | Terminal-Bench-Science 0.1 (70 items), CritPt v2 Hard (40 items) | 3, 5 | 15%, 10% |
| General | AA-Omniscience v1.1 (3,000 items), HLE-Diamond (1,000 items) | 1, 1 | 15%, 10% |
The evaluation follows standardized, zero-shot instruction prompting and consistent scoring principles. The methodology specifies a temperature of 0 for non-reasoning models and 0.6 for reasoning models; API failures are automatically retried up to 30 times. The suite is primarily English-language text; some evaluations, such as GDP.pdf and HLE-Diamond, include image inputs.
Hydrogen xhigh: 54 points, 14.7 minutes per index task, and an average cost of $3.16 per task.
Hydrogen high: 51 points, 8.3 minutes per task, and $1.87 per task.
Hydrogen medium: 41 points, 3.8 minutes per task, and $0.80 per task.
Hydrogen low: 27 points, 1.4 minutes per task, and $0.29 per task.
The X post adds that xhigh scores 3 points above high but uses nearly twice as many output tokens; the model page reports that xhigh generated 304 million output tokens on the index evaluations.
Gemini 4 Hydrogen model page: Includes scores, costs, runtimes, and token totals for each reasoning tier.
Intelligence Index evaluation methodology: Includes evaluation components, item counts, repetitions, weights, and general parameters.
The results support the conclusion that xhigh scores higher overall than the model's other reasoning tiers, but the gain comes with longer runtimes and higher costs. Artificial Analysis explicitly notes that the index is an aggregate capability metric and does not map directly to every use case. The materials for this evaluation do not disclose the model endpoint, exact model identifier, or per-item results, so these pages alone are not enough to rerun the same tests. The API availability status shown on the Artificial Analysis model page at the time did not match the Google API release notes; Google's documentation should be treated as authoritative for the API's public status. The page data is a dynamic snapshot collected on 2026-11-16.
The original post, model page, and methodology page allow readers to verify the published aggregate results and general settings. Rerunning the individual tests would still require the same model endpoint, prompts, dataset, and complete evaluation artifacts.
Gemini 4 Hydrogen