Gemini 4 Hydrogen Focuses on Speed and Cost in Agentic Coding
Google launches Gemini 4 Hydrogen, already available in Gemini, Search, AI Studio, and Antigravity. The model targets fast agentic tasks and shows speed gains over Gemini 3.8 Flash, with a standard rate of $1.50 per million input tokens.
Google launched Gemini 4 Hydrogen on November 12, 2026, a model designed for fast agentic tasks, coding, and document processing. It is already available in Gemini, Google Search, AI Studio, and Antigravity, with a clear positioning: to be fast, cost less than the heaviest models, and perform well on everyday digital tasks. On some benchmarks, it even approaches Gemini 4 Argon, without exactly playing in the same category. The cost, however, remains lighter than that of cutting-edge models — which quickly matters when tokens pile up.
Google vise le travail quotidien, pas seulement les gros calculs
Hydrogen is presented as the “daily driver” of the Gemini 4 family, while Carbon and Argon handle deeper reasoning tasks. This role separation says a lot about Google's strategy: reserving the most expensive models for the most demanding cases, and pushing Hydrogen for uses where speed and fluidity are paramount. Koray Kavukcuoglu, now head of Google DeepMind, sums it up by calling it a model “built for rapid execution.”
The model is now offered to free users of the Gemini app, but also via the Gemini API, Google AI Studio, Gemini Enterprise, and, for Google AI Pro or Ultra subscribers, in Antigravity. Google Search also uses it worldwide. In other words, Hydrogen is not confined to a developer sandbox; it becomes a production component. This is where things get serious, and a bit less theoretical.
Des gains mesurés sur le code et les agents
On the FrontierCode 2.0 benchmark, which evaluates real engineering tasks based on quality and “mergeability,” Hydrogen scores 38.2%. The score remains close to GPT-6.1 Sol at 39.0%, and Gemini 4 Carbon at 40.0%. But the point that changes the interpretation of the result is the cost: $2.40 per task, compared to $4.04 for Carbon, about 40% less. At average reasoning effort, Google even says Hydrogen outperforms the best result of Gemini 4 Argon at less than a seventh of the cost per task.
On Terminal-Bench 4.1, which measures terminal work, Google announces a 57% success rate for Hydrogen, compared to 18.2% for Gemini 3.8 Flash. The model still lags behind Claude Sonnet 5.5, at 63.9% in Google's table. In feedback shared by Google and early testers, Hydrogen also improves recovery after a failed tool call or an empty search: it retries, switches files, or tries another approach instead of stopping abruptly. It's less spectacular than a grand speech about AI, but much more useful daily.
Un modèle plus rapide, mais pas toujours meilleur partout
Google claims Hydrogen executes tasks 2.5 times faster than Gemini 3.8 Flash at its default reasoning setting. In independent tests cited by VentureBeat, it scores 51 on the Artificial Analysis Intelligence Index, versus 31 for Flash at high reasoning level, and one point ahead of Gemini 4 Argon at high level. With the new xhigh reasoning level, it reaches 54 according to Google.
The gains are not uniform. On some coding tests, Hydrogen climbs significantly; on others, the gap narrows. On Terminal-Bench 4.0, Google indicates 54% at both high and xhigh, with no additional gain from the highest setting. For teams already paying for inference, this matters: more reasoning does not always mean better performance. The model is thus less a bulldozer than a well-sharpened pocket knife.
Ce que ça change pour les développeurs et les entreprises
The switch from Gemini 3.8 Flash to Hydrogen is not just about the model name. Migration documentation signals an important behavior change: Hydrogen starts at high reasoning by default, versus medium on Flash, and adds an xhigh option. The old thinking_budget parameter is no longer accepted, nor is the MINIMAL mode. Google also changes cache management: on Google Cloud, automatic caching starts at 4,096 tokens, versus 6,144 on Flash, and explicit cache is not available at launch.
The model accepts text, images, audio, and video, and Google Cloud can directly ingest Office and OpenDocument files. Its context of about 1 million tokens is shared between input, output, reasoning tokens, and tool calls. In practice, this makes it suitable for uses like reading tables, long documents, multimodal reports, and assistants capable of chaining several actions without losing track. For a company, the real question is not just “does it work?”, but “does it avoid rework, errors, or hidden costs?”.
Le prix reste central dans la bataille entre modèles
Google lists a standard price for Hydrogen of $1.50 per million input tokens and $7.50 per million output tokens, excluding cache, with identical rates across the entire context window. This places the model at the level of the future standard price of Flash, but still above the cheapest market offers. Google also offers Batch, Flex, and Contributor modes at half price, under specific conditions of delay, priority, or data sharing.
Compared to Claude Haiku 5.5, GPT-6 Luna, or Muse Ember, Hydrogen sits in an intermediate zone: more expensive than entry-level models, but more performant on several agentic and coding tasks. On the Artificial Analysis Intelligence Index, the table published by VentureBeat places Hydrogen at 51 at high reasoning level, with an estimated cost of $1.87 to $3.16 per task depending on the setting. Chinese open-weight models like DeepSeek V4.1 Flash 1017 or Xiaomi MiMo-V2.6-Flash remain very aggressive on price. The trade-off thus becomes very concrete: capacity, latency, cost, and sometimes deployment sovereignty.
Key points
- Hydrogen has been available since November 12, 2026.
- Google deploys it in Gemini, Search, AI Studio, and Antigravity.
- FrontierCode 2.0: 38.2% for Hydrogen, 40.0% for Carbon.
- Terminal-Bench 4.1: 57% for Hydrogen versus 18.2% for Flash.
- Standard price: $1.50 input and $7.50 output per million.
- Google says Hydrogen is “built for rapid execution.”
By the numbers
- 2.5 times — announced speed versus Gemini 3.8 Flash, default setting, Google.
- 1 million tokens — approximate context window, Google Cloud, 2026 launch.
- 57% — Hydrogen's score on Terminal-Bench 4.1, Google, November 2026.
- $2.40 per task — average cost on FrontierCode 2.0, source Google / published data.
- 81% — announced reduction in “overreach” cases in agentic tasks, Google.