Google has released Gemini 4 Hydrogen, a model designed for coding and agentic workflows. It serves as a daily driver for everyday agent loops, positioned below the company’s frontier models, Gemini 4 Carbon and Gemini 4 Argon. The model is closed-source and runs exclusively on Google-hosted APIs. Self-hosting is not supported, and no weights or quantized variants are available.
What is Gemini 4 Hydrogen?
Hydrogen adapts the architecture of Gemini 4 Carbon. Google does not disclose parameter counts, specific architecture details, or training hardware. The knowledge cutoff stands at March 2026. Input accepts text, images, audio, video, PDFs, tables and charts. Output is text only. The unified window handles input, thinking and output tokens at 1,048,576 tokens.
How do the 4 thinking levels work?
Developers configure the thinking_level within the generation_config. Options include low, medium, high and a new xhigh. The default setting is now high, replacing the medium default found in Gemini 3.8 Flash. A minimal setting now triggers an error. The model card warns that higher levels consume more tokens.
How does Hydrogen work with Carbon and Argon?
The design relies on role-splitting across models. In Google Antigravity, an Orchestrator option switches models within a single task. Carbon handles exploration, ambiguity resolution and planning. Hydrogen edits files, updates packages and runs tests. When Hydrogen encounters friction, control returns to Carbon. On Terminal-Bench 4.1, this pairing achieved near-Carbon task success rates. It used 24% less quota and finished 19% faster than Carbon alone. The Antigravity SDK exposes this split for custom multi-agent systems.
How does Gemini 4 Hydrogen perform on benchmarks?
All figures are self-reported by Google unless noted.
- Artificial Analysis Intelligence Index: Scores 51 at high and 54 at xhigh. Argon scores 50 on Index v5.0.
- Terminal-Bench 4.1: Achieves 57%, compared to 18.2% for Gemini 3.8 Flash.
- HLE-Diamond: Scores 58.8%, leading its class.
- Vals Tax Agent Bench: Scores 52.3%, ranked #1 overall.
- Vals Legal Research Bench: Scores 55.8%, tied for #1.
- Multimodal Knowledge Agent: Scores 53.4%.
- Agents’ Last Exam v1.1: Scores 71.6%.
- Agent Arena: Ranks #5.
Speed is the headline claim. Hydrogen finishes tasks 2.5x faster than Gemini 3.8 Flash at the default high level. In partner tests, it ran 2x to 2.8x faster with 12% to 23% fewer tokens. Google reports more than 5x fewer fabricated claims than Gemini 3.8 Flash. It scored 96.7% on an internal agent honesty evaluation. It was 2.3x more resilient to indirect prompt injection on Gray Swan’s benchmark.
How much does Gemini 4 Hydrogen cost?
The API price is $1.50 per 1M input tokens and $7.50 per 1M output tokens. Pricing is flat across the full window. Output pricing includes thinking tokens. Implicit caching cuts input to $0.075 per 1M for prompts of 4,096 tokens or more. Explicit caching is not supported.
Context matters for the price. Gemini 3.8 Flash costs $0.75/$3.75 until December 31, 2026. From January 1, 2027, it rises to $1.50/$7.50, matching Hydrogen. Argon costs $2/$10 until November 30, 2026, then $4/$20.
Where can developers use Gemini 4 Hydrogen?
The model ID is gemini-4-hydrogen. It is live in the Gemini API, Google AI Studio on all tiers, and Android Studio. It also ships in Gemini for business, the Gemini app, Google Search and Sheets. Antigravity access requires Google AI Pro or Ultra. Vertex AI is not mentioned in the launch posts.
Migration from 3.8 Flash needs a few changes. Remove temperature, top_p and top_k, since the server ignores them. Remove cached_content fields. Raise hardcoded max_output_tokens caps to 65,536.
Gemini 4 Hydrogen vs competitors
| Feature | Gemini 4 Hydrogen | Gemini 3.8 Flash | Gemini 4 Argon | Claude Haiku 5.5 |
|---|---|---|---|---|
| Release | Nov 12, 2026 | Sep 2, 2026 | Oct 13, 2026 | Oct 7, 2026 |
| Parameters | Not disclosed | Not disclosed | Not disclosed | Not disclosed |
| Context | 1,048,576 unified | Not stated in launch post | 1M unified | Not stated on launch page |
| Weights | Closed | Closed | Closed | Closed |
| Input $/1M | $1.50 | $0.75 (intro), $1.50 from Jan 1, 2027 | $2 (intro), $4 from Dec 1, 2026 | $0.10 (≤100K), $0.50 (>100K) |
| Output $/1M | $7.50 | $3.75 (intro), $7.50 from Jan 1, 2027 | $10 (intro), $20 from Dec 1, 2026 | $0.50 (≤100K), $2.50 (>100K) |
| AA Intelligence Index | 51 (high), 54 (xhigh), per Google | Not disclosed | 50 (v5.0) | 36 (v5.0.1) |
| Terminal-Bench | 57% (4.1) | 18.2% (4.1, per Google) | 57% (4.0, AA) | 39.2% (4.0, max effort) |
| Reasoning control | 4 thinking levels | Effort levels | Not stated | Effort settings |
Terminal-Bench 4.0 and 4.1 are different versions, so cross-version scores are not directly comparable.
What it means
Developers building coding agents face a distinct trade-off. Hydrogen offers superior speed and terminal command accuracy compared to the previous Flash model, but it is strictly a closed API service. You cannot run it on your own hardware. The model also enforces a hard cap on output tokens at 65,536, which limits long-form generation in a single response. Pricing has aligned with the upcoming standard rates for the Flash model, removing the introductory discount. For complex tasks, the recommended approach is to pair Hydrogen with Carbon in the Antigravity framework, which reduces token usage and increases task completion speed.
