Summary:
Gemini 4 Hydrogen targets faster agent execution. Its launch scorecard shows knowledge-work gains and terminal-task trade-offs, with important context-window and benchmark-method details.Google announced Gemini 4 Hydrogen on November 12, 2026, positioning it as the faster execution model alongside Carbon and Argon for deeper, longer-running work. Rollout begins across the Gemini API, AI Studio, Gemini for business, the Gemini app and Search. Antigravity access starts with paid Pro and Ultra plans.
Availability and API pricing
Google lists $1.50 per million input tokens and $7.50 per million output tokens, with the same rates across the context window. In the free Gemini app, Hydrogen becomes available through automatic model selection when appropriate. That is different from promising every free user a manual Hydrogen selector.
Benchmark comparison: Hydrogen versus Sonnet 5.5 and GPT-6.1 Terra
The scorecard below reproduces selected numerical results from Google's November 2026 evaluation report. Higher is better in every row; bold marks the best reported numeric result among these three models. A missing result is not a zero. These are different tests, not components of a combined ranking.
| Benchmark | Gemini 4 Hydrogen | Claude Sonnet 5.5 | GPT-6.1 Terra |
|---|---|---|---|
| AutomationBench-AA | 76.1% | 71.8% | 60.2% |
| HLE-Diamond | 58.8% | 53.3% | 47.3% |
| Vending-Bench 2 (bank balance) | $14,549 | $11,165 | $9,107 |
| Terminal-Bench 4.1 (cheating-adjusted) | 57.0% | 63.9% | Not reported |
| OSWorld 2.1 (82-task offline subset) | 77.6% | 83.9% | 74.2% |
| Terminal-Bench Science 0.1 | 51.4% | 59.9% | 58.3% |
How the comparisons were measured
For AutomationBench-AA and HLE-Diamond, Google cites Artificial Analysis; for Vending-Bench 2, it cites Andon Labs. Those are independent leaderboard sources as compiled in Google's launch report, rather than a new unified test conducted by Google.
The remaining rows mix developer measurements and published peer results. Hydrogen's Terminal-Bench runs use a mini-swe-agent vision harness, averaging five attempts per task; Terminal-Bench 4.1 uses high effort. The OSWorld comparison uses the 82-task offline subset, not the full benchmark: Google reports its Hydrogen and Sonnet measurements as the maximum over three single-attempt runs, while Terra comes from its system card. Google's general policy is highest available thinking settings, with best available peer results when those settings are not reported. Harnesses and sampling differ, so the table should guide task-specific testing, not establish universal model superiority.
The million-token limit needs a qualification
The model card specifies a combined input-output context window of up to one million tokens. Synchronous output defaults to 64,000 tokens; background long decode can reach one million. An application should not budget one million input tokens plus another million output tokens in the same context.
What to try first
For an existing agent workflow, compare completion quality, total token cost and elapsed time on the same tasks before switching models. The launch results favor Hydrogen on several knowledge-work measures, while Sonnet leads the terminal and offline computer-use rows shown here. Choose the test that resembles the work, rather than treating the largest percentage as the most relevant one.
Sources and image credits
- Google: Introducing Gemini 4 Hydrogen, November 12, 2026
- Google DeepMind: Hydrogen evaluation approach, methodology and results
- Gemini 4 Hydrogen model card
- Original scorecard composed from the cited numerical results. The official Gemini family spark identifies the model family; it is not a distinct Hydrogen logo. Logo Source: Google, from the Google Image Library, which permits publication with credit.
Google and Gemini are trademarks of Google LLC. This article is not endorsed by or affiliated with Google.
