Google launches Gemini 4 Hydrogen, a faster, lower-cost model for coding and agent workflows

Google on 12 November introduced Gemini 4 Hydrogen, a faster, lower-cost Gemini 4 model for everyday coding, agent-based (also called agentic) workflows and multimodal knowledge work.

Google said Hydrogen will be available across Google AI Studio, the Gemini API, Gemini for Business, the Gemini app, Google Search and in paid Google AI Pro and Ultra tiers inside Google Antigravity. The model supports a 1 million-token input context (tokens are units of text) and can generate up to 1 million output tokens.

Pricing and performance: Google lists a flat price across Hydrogen's full 1 million-token context window of $1.50 per million input tokens and $7.50 per million output tokens. It says Hydrogen surpasses Argon on the Artificial Analysis AI Index at the default high thinking level (51 vs. 50, reaching 54 at xhigh).

Google also published benchmark results it says show real-world speed and cost gains: in Orchestrator (Google Antigravity) average quota consumption fell by 24% and task completion ran 19% faster on Terminal-Bench 4.0. Compared with Gemini 3.8 Flash, Hydrogen finished tasks about 2.5x faster at the default high setting, improved Terminal-Bench 4.1 scores (57% vs. 18.2%) and reduced fabricated claims by more than 5x. Those figures are reported by Google from its benchmark runs.

Tools and use cases: Google positions Hydrogen for iterative coding and debugging, multimodal synthesis of PDFs, spreadsheets, audio and video, and agent-based work such as running parallel worker and verifier agents. New tooling examples include Orchestrator (dynamic model switching) and Antigravity demos that show Hydrogen powering Sheets automations, front-end design and autonomous audit-and-fix loops.

What this means for African developers and enterprises: Hydrogen's pricing and its 1 million-token context affect budgets for long-document or long-code generation because output tokens cost more than input tokens. Teams that process large codebases, long reports or long-form outputs should plan budgets around the output token rates and expect the model to generate many output tokens for extensive tasks.

Where requests are processed also matters. Google Cloud documentation notes that models are served from regional, multi-region and global endpoints and that TPU and AI-zone availability varies by region and zone. Choosing a global endpoint can improve availability but means you cannot control which region actually processes requests; regional or multi-region endpoints exist for customers that need processing to stay within jurisdictional boundaries. TPU and AI-zone hardware needed for large-scale workloads is not available uniformly across all regions.

For African customers this matters in three ways: latency (requests routed to distant regions add delay), cost (cross-region bandwidth and longer-running jobs can raise bills) and compliance or data-residency rules (using a global endpoint may not meet jurisdictional processing requirements). Organisations that require in-region processing or dedicated accelerators should check Google Cloud's region and AI-zone maps and select regional or multi-region endpoints where they are offered.

Google describes Hydrogen as a model built for fast, disciplined execution in everyday developer workflows and as a cost-efficient worker in multi-model setups that also use Carbon or Argon for deeper reasoning.

Newzroom dot Africa has reached out to Google and Google Cloud for comment and is awaiting a response. This story will be updated when they respond.

Source: Google DeepMind