Google DeepMind's model card reports Gemini 4 Hydrogen's scores across several benchmark categories and compares them with Claude Sonnet 5.5 and GPT-6.1 Terra. These are aggregate figures published by the model provider; the page does not provide enough configuration details or per-item records to independently rerun each test.
Tasks this can help assess: Understand Google DeepMind's account of Hydrogen's capability range, published benchmark results, and disclosed limitations.
Tasks this should not be extrapolated to: Predict performance across all real-world tasks or across different clients or parameter settings, or directly compare its scores with rankings from Arena, Artificial Analysis, or Plotly.
Applicable model version: Gemini 4 Hydrogen; the page does not specify an API model ID, revision, or the reasoning level used for each test.
Test environment or client: Not stated; the full execution environment for individual benchmarks is not included in the model card.
Reasoning level and parameters: Not stated; the page only says the product can adjust speed and intelligence across four thinking levels.
The model card says its evaluation covers agentic coding, knowledge work, computer use, multimodal understanding, science and math, and reasoning. The results table gives notes on specific scopes or scoring methods: for example, Terminal-Bench 4.1 uses a cheating-adjusted score, OSWorld-2.1 uses an offline subset, MKA reports Climb WSR accuracy, and Vending-Bench 2 reports account balance. The model card links to a separate evaluation-methods page, but does not detail the benchmark sets, prompts, tool configurations, number of runs, statistical intervals, or scoring details on this page.
The figures below should therefore be treated as Google DeepMind's official reported values as of November 2026. Although the model card lists comparison models, its results table does not specify the endpoint versions and parameters for each model or whether testing used identical configurations. The table alone cannot establish that differences are due to model capability itself.
| Category | Benchmark | Notes | Gemini 4 Hydrogen | Claude Sonnet 5.5 | GPT-6.1 Terra |
|---|---|---|---|---|---|
| Agentic coding | SWE-sweep | — | 13.5% | 3.1% | 5.8% |
| Agentic coding | Terminal-Bench 4.1 | Cheating-adjusted | 57.0% | 63.9% | — |
| Knowledge work | Agents' Last Exam v1.1 | Score | 71.6% | 60.4% | 69.8% |
| Knowledge work | AutomationBench-AA | — | 76.1% | 71.8% | 60.2% |
| Knowledge work | Vending-Bench 2 | Bank balance (USD) | $14,549 | $11,165 | $9,107 |
| Computer use | OSWorld-2.1 | Offline subset | 77.6% | 83.9% | 74.2% |
| Multimodal understanding | MKA | Climb WSR accuracy | 53.4% | 52.0% | 43.8% |
| Multimodal understanding | Chartography | — | 66.6% | 61.6% | 63.2% |
| Science and math | Terminal-Bench-Science 0.1 | — | 51.4% | 59.9% | 58.3% |
| Reasoning | HLE-Diamond | — | 58.8% | 53.3% | 47.3% |
The model card also reports percentage-point changes in internal automated safety evaluations relative to Gemini 3.8 Flash: text safety +0.80pp, multilingual safety +1.10pp, image-input safety +7.00pp, refusal tone -0.65pp, and unjustified refusals +5.57pp. The card says higher is better for the first three safety metrics and for refusal tone, while lower is better for unjustified refusals; the last result is therefore a regression. These were automated evaluations, not human evaluations or red-team tests. Google says the evaluation sets and methods have been updated, so these results cannot be directly compared with earlier Gemini model cards.
Model positioning: The card says Hydrogen is adapted from Gemini 4 Carbon and is positioned for fast execution in everyday coding and agentic workflows.
Inputs and context: Supports text, images, audio, and video; the combined input and output context window is up to 1M tokens.
Output: Up to 64K tokens by default in synchronous output; up to 1M tokens with background long decode.
Knowledge cutoff: March 2026.
Known limitations: May hallucinate; may have latency or timeouts; may use more tokens at higher thinking levels; some domain knowledge may not be current enough.
Safety disclosure status: The full Safety & Security Report was still marked as coming soon at the time of collection. The model card says Hydrogen did not reach its CBRN and cybersecurity CCLs, but did reach the warning thresholds for both CCLs, and says mitigations continue to be deployed. This records the card's conclusions without drawing further inferences about risk capabilities.
The official results show a mix of strengths and weaknesses across the listed benchmarks. For example, Hydrogen's reported values are higher than at least one of the two comparison models on SWE-sweep, Agents' Last Exam, AutomationBench-AA, Vending-Bench 2, and HLE-Diamond; on Terminal-Bench 4.1, OSWorld-2.1, and Terminal-Bench-Science 0.1, it scores below Claude Sonnet 5.5. These results are more useful for showing which tasks and figures Google chose to publish than for concluding that Hydrogen leads across agentic tasks in general.
These scores come from the model provider's model card. They are not user-preference rankings from Arena, a composite index from Artificial Analysis, or a one-off data analytics exam from Plotly. Those sources use different tasks, harnesses, and aggregation methods, so their results should not be combined or compared directly. The model card page may change; this article records the snapshot visible on 2026-11-16.
The model card text is enough to verify the table's figures and notes, but not to rerun each test. Exact reproduction would also require each benchmark's version and test set, model endpoint and revision, prompts, tools/harness, reasoning level, sampling parameters, number of runs, and per-item results. The card links to an evaluation-methods page, but this article relies only on the model card and does not treat details absent from it as evidence.
Gemini 4 Hydrogen