Gemini 4 Hydrogen: $1.50 per input token, faster execution

What changed

Google announced Gemini 4 Hydrogen on November 12, 2026, a new model designed for fast execution across everyday coding and agentic workflows. The company made Hydrogen available across all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro and Ultra subscribers in Google Antigravity.

Hydrogen carries flat pricing of $1.50 per million input tokens and $7.50 per million output tokens across its entire 1 million-token context window. On the Artificial Analysis Intelligence Index at its default high thinking level, Hydrogen scores 51 compared to Argon’s 50, and reaches 54 at its highest xhigh thinking setting. The model operates at four thinking levels, low, medium, high (default), and xhigh, allowing teams to adjust reasoning depth against latency and cost.

Google also introduced Orchestrator, a new feature in Google Antigravity that automatically switches between frontier models like Carbon and Hydrogen within the same session, eliminating the need to select a single model at the start. This removes friction from multi-model deployments by routing tasks to the most appropriate model as work progresses.

Why it matters

Hydrogen addresses a practical constraint in AI workflows: frontier models excel at deep architectural reasoning but waste tokens on straightforward execution tasks. Most organizations face a choice between using capable models everywhere and accepting higher costs and latency, or manually deciding which tasks need which model. Orchestrator automates that routing.

According to Google’s testing on Terminal-Bench 4.0, Orchestrator achieved near-Carbon task success rates while consuming 24 percent less quota and completing tasks 19 percent faster on average. For teams building complex systems, this efficiency matters. Carbon can coordinate overall strategy while Hydrogen powers worker and verifier agents to build and test changes simultaneously across isolated workspaces, compressing development cycles without inflating token costs.

For developers migrating from Gemini 3.8 Flash, Hydrogen benchmarks show material gains. The model finishes tasks 2.5 times faster at its high default setting while more than tripling Terminal-Bench 4.1 scores (from 18.2 percent to 57 percent). Early partner testing across complex graphs, SQL, and data-heavy workflows showed Hydrogen running 2x to 2.8x faster than 3.8 Flash while reducing token usage by 12 to 23 percent. The company reports fabricated claims decreased by more than 5x compared to 3.8 Flash.

Hydrogen also ranks favorably in open-ended agentic work. On Agent Arena, a live benchmark of agentic execution, Hydrogen ranks 5th. For domain-specific tasks, it leads its class on HLE-Diamond (58.8 percent), ranks 1st overall on Vals Tax Agent Bench (52.3 percent), ties for 1st on Vals Legal Research Bench (55.8 percent), and leads its class on AutomationBench-AA.

For multimodal knowledge work, Hydrogen supports native 1 million-token input paired with up to 1 million output tokens. It scores 53.4 percent on the Multimodal Knowledge Agent benchmark for end-to-end agentic knowledge work over complex visual files and 71.6 percent on Agents’ Last Exam v1.1 for professional workflows. These benchmarks are new with limited public track records, making real-world validation important before production deployment.

What to test

Before adopting Hydrogen across workflows, verify these specific claims in your environment:

• Orchestrator’s 24 percent quota savings and 19 percent speed gains. These come from Terminal-Bench 4.0, a single benchmark. Test Orchestrator against your actual workflows, not public benchmarks. Measure quota usage and latency in production before scaling.

• 2x to 2.8x faster execution on repository tasks. Google attributes this to parallel file-read and search calls rather than sequential file inspection. Verify this claim with your own codebase. Repository complexity varies widely; early partner results may not replicate in your stack.

• More than 5x reduction in fabricated claims compared to 3.8 Flash. Factual accuracy is critical for agentic work. Run hallucination tests on tasks relevant to your domain. A five-fold improvement is significant; confirm it holds for your use cases before production deployment.

• Ranking 5th on Agent Arena for open-ended agentic execution. Arena.ai’s live benchmark reflects real-time model performance. Check where Hydrogen ranks in your testing window. Benchmark scores shift as vendors release updates.

• The four thinking levels offer smooth progression across quality, latency, and cost. Test each level against representative tasks to understand actual tradeoffs. Smoothness is subjective; your optimization target may differ from Google’s.

• Hydrogen’s multimodal performance on dense PDFs, spreadsheets, and video. The company reports 53.4 percent on the Multimodal Knowledge Agent benchmark and 71.6 percent on Agents’ Last Exam v1.1. These are new benchmarks with limited track records. Test on your document types before using it for production knowledge work.

Run internal benchmarks on your own code repositories, domain-specific datasets, and agentic workflows. Vendor benchmarks prove capability but rarely predict performance in your specific stack.

Common questions

When is Gemini 4 Hydrogen available?

Gemini 4 Hydrogen launched November 12, 2026. The model is available across Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro and Ultra subscribers in Google Antigravity.

How does Gemini 4 Hydrogen pricing compare to other models?

Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens. This undercuts frontier models like Carbon and Argon. Compare these rates directly against your current vendor contracts to evaluate total cost impact.

What is Orchestrator, and when can I use it?

Orchestrator is a feature in Google Antigravity that automatically switches between models like Carbon and Hydrogen within the same session. According to Google, it achieved near-Carbon task success rates while consuming 24 percent less quota and completing tasks 19 percent faster on Terminal-Bench 4.0. It is available now in Google Antigravity, but real-world performance depends on your workflows.

The conclusion

Hydrogen addresses a genuine gap in the model landscape. Frontier models are powerful but expensive; smaller models are fast but often too limited for complex work. Hydrogen targets the middle ground, and Orchestrator removes friction from multi-model workflows by automating model selection. For teams already invested in Google’s AI stack, the pricing and benchmark claims warrant testing.

The real validation occurs in production. Terminal-Bench and Agent Arena scores are legitimate, but your workflows differ. Run internal benchmarks on your codebase, domain data, and agentic tasks before migrating budget from your current vendor. If Orchestrator’s 24 percent quota savings and 19 percent speed gains hold in your environment, Hydrogen becomes a practical cost control mechanism. If they don’t, you’ll need to reassess manual model selection or deeper frontier model reliance.

Monitor vendor benchmark updates as Hydrogen competes openly with other models in its class. Watch also for real-world adoption data. Early partner results are encouraging, but adoption at scale often reveals edge cases vendor benchmarks miss.

Related reading

• Gemini Decisions API: $0.10 per million input tokens
• Gemini 4 Carbon vs GPT-6.1 Sol: Pricing and Performance
• Gemini Omni 1.2 Flash: 30% price cut, 15-second video