Google released Gemini 4 Hydrogen on November 12, 2026. In its announcement, Google positions the model as a fast, cost-efficient option for everyday coding and agentic work, with its larger Gemini 4 models, Carbon and Argon, handling deep, long-horizon reasoning. It is described as a standalone upgrade for teams on Gemini 3.8 Flash. All benchmark figures below are reported by Google unless stated.
GoogleWhat shipped
- Pricing. $1.50 per million input tokens and $7.50 per million output tokens, flat across the full 1M-token context window, per Google.
- Limits. Up to 1M input tokens and up to 1M output tokens.
- Thinking levels. Four: low, medium, high (the default) and a new xhigh.
- Availability. Gemini API and Google AI Studio on all tiers, Android Studio, and Google Antigravity for paid AI Pro and Ultra subscribers. Enterprises get it through Gemini for business and Gemini in Google Sheets for eligible subscriptions. Consumers get it in the Gemini app model dropdown, in Google Search for all users globally, and, for free users, through automatic model selection when appropriate.
The benchmarks, read carefully
| Benchmark | Gemini 4 Hydrogen (Google-reported) | Context |
|---|---|---|
| Artificial Analysis Intelligence Index | 51 at high, 54 at xhigh | Argon is cited at 50 |
| Terminal-Bench 4.1 | 57% | Gemini 3.8 Flash: 18.2% |
| HLE-Diamond | 58.8% | Described by Google as leading its class |
| Vals Tax Agent Bench | 52.3% | Ranked first overall, per Google |
| Vals Legal Research Bench | 55.8% | Tied for first, per Google |
| Multimodal Knowledge Agent | 53.4% | Described as ahead of other models |
| Agents' Last Exam v1.1 | 71.6% | No comparison given |
| Agent Arena (arena.ai) | Ranked fifth | Rank, not a score |
Terminal-Bench 4.1: Gemini 4 Hydrogen vs Gemini 3.8 Flash
Company-reported success rate, in percent, on an agentic terminal benchmark.
Vendor-reported figures; results depend on thinking level and harness settings.
Two cautions apply. First, the headline jump on Terminal-Bench 4.1 (18.2% to 57%) is a company-reported comparison against the previous Flash tier, so it measures generation-over-generation progress, not position against other vendors' models. Second, the Intelligence Index figure comes from a third-party index, but Google's post is the source for how it is presented, so it is worth checking the index directly. The benchmarks list also includes AutomationBench-AA and SWE-sweep, where Google gives no exact scores.
Speed and efficiency claims
- Finishes tasks about 2.5 times faster than Gemini 3.8 Flash at its default setting.
- Early partners in graph, SQL and data-heavy enterprise work reported 2x to 2.8x faster runs and 12% to 23% lower token usage versus 3.8 Flash.
- Google Sheets canvas outputs are over 30% faster than previous models.
- In Google Antigravity, an orchestrator that switches between Carbon and Hydrogen in one session reached near-Carbon task success on Terminal-Bench 4.0 while using 24% less quota and finishing 19% faster on average.
The last item is the most analytically interesting, because it describes a design pattern rather than a single-model result: a stronger model plans, and a cheaper one executes. Google says that in multi-model workflows Carbon can coordinate while Hydrogen runs worker and verifier agents in parallel.
Pricing arithmetic
Because the rate is flat across the whole context window, cost scales linearly with tokens and does not step up for long prompts. Our illustrative examples, with assumed token counts:
| Workload (assumed) | Input tokens | Output tokens | Cost |
|---|---|---|---|
| Short coding turn | 20,000 | 2,000 | $0.03 + $0.015 = $0.045 |
| Large repository pass | 200,000 | 20,000 | $0.30 + $0.15 = $0.45 |
| Full 1M-token context read | 1,000,000 | 10,000 | $1.50 + $0.075 = $1.575 |
The token counts are our assumptions, not Google's. Two forces pull in opposite directions in agent workloads: lower per-token price, and, per Google's partner results, 12% to 23% fewer tokens used. Higher thinking levels such as xhigh would be expected to spend more output tokens, so real cost depends on the level chosen.
Safety and reliability claims
Google reports several vendor-measured improvements versus 3.8 Flash. These are internal or partner evaluations and have not been independently reproduced in the sources we reviewed:
- Reduces agentic overreach by up to 81% and under-service by up to 79%.
- Cuts misaligned actions and task shortcutting 12x in tool-use and coding evaluations.
- Up to 2x less likely to bypass permission boundaries when tasks are blocked.
- 96.7% honesty rate on an internal evaluation, and 99.9% chain-of-thought legibility across more than 200,000 audited turns.
- More than 2.3x more resilient to indirect prompt injection on the Gray Swan IPI benchmark.
- False refusals on HealthBench Pro under 0.4%.
What this means for teams
- Tiered model use is becoming the default design. Pairing a frontier planner with a fast worker is now described by a model vendor as a first-class workflow, not a workaround.
- Flat long-context pricing simplifies budgeting. There is no pricing cliff at long prompts, though quality at very long context should be tested on your own data.
- Benchmarks need your own data. Vendor numbers differ by harness and thinking level. Running the same task set across models is the reliable comparison, which is the use case for multi-model workspaces like Metir that let teams switch models per task.
Sources:
- Google: Introducing Gemini 4 Hydrogen
- The Neuron: Everything that happened in AI today, Friday, November 13
Image credits
- Aerial view of the Google campus, Mountain View, California: Austin McKinley, CC BY 3.0, via Wikimedia Commons. A file photo of Google's campus, not of the model.