SummaryGoogle released Gemini 4 Hydrogen on November 12, 2026. This free-tier model is built on the Gemini 4 Carbon infrastructure. It claims to operate at 2.5 times the speed of Gemini 3.8 Flash and supports inputs up to 1 million tokens. It targets agentic coding and multi-step workflows. Additionally, it arrives with a new 'Orchestrator' feature in Google Antigravity. This feature automatically routes tasks between Carbon and Hydrogen, reducing usage by approximately 24% and time by approximately 19%. |
Why it mattersBy bringing a near-frontier reasoning model to free Gemini app users and all Google Search users, Google is narrowing the cost-intelligence gap that has largely kept agentic AI as a paid developer tool. The Orchestrator's automatic model switching signals a shift from a 'pick one model' approach to a managed multi-model pipeline. This could reshape how teams budget for LLM inference. A rapid release cadence of three new Gemini 4 variants in six weeks is forcing Anthropic and OpenAI to match them on both speed and price. |
Highlights- SWE-sweep score of 13.5% (Claude Sonnet 5.5: 3.1%, GPT-6.1 Terra: 5.8%) (Google data)
- Artificial Analysis positions general intelligence just below Sonnet 5.5 / GPT-6.1 Sol. The cost is about ⅓ of Sonnet's, but about 2.5 times that of GPT-6.1 Sol.
- The Orchestrator's automatic switching between Carbon and Hydrogen reduces quota consumption by 24% and task duration by 19% in Google's internal tests.
|
Critical takeNearly all the figures in the headline—2.5x speed, 57% Terminal-Bench, 13.5% SWE-sweep—come from Google's own blog or its preferred benchmark suite. The 'free' framing ignores that the Orchestrator workflow, where real efficiency gains occur, is locked behind AI Pro/Ultra subscriptions. Three Gemini 4 releases in six weeks raises the question of whether each update represents a genuine capability leap or merely a gradual re-tuning of the same Carbon backbone. |
This critique was written by the local AI model
Qwen3.8-27B. I chose the story; I do not review each critique before it is published.