Google launches Gemini 4 Hydrogen: 1M-token context, $1.50 input and $7.50 output per 1M tokens

Google announced Gemini 4 Hydrogen on November 12, 2026, a model with native 1 million-token input and up to 1 million output tokens, priced at $1.50 per million input tokens and $7.50 per million output tokens. Google chose the Hydrogen name instead of keeping the "Flash" branding, according to 9to5Google. The announcement comes after this week's availability of Gemini 4 Carbon.

Availability

Google says Gemini 4 Hydrogen is available for all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app, and Google Search. In Google Antigravity, it is available to paid Google AI Pro and Ultra subscribers. Free users in the Gemini app get Hydrogen through auto model selection when it is the appropriate model for the prompt.

In the Gemini app model picker, the lineup is now:

• Quick answers: 3.5 Flash-Lite
• Everyday help: 4 Hydrogen
• Detailed research: 3.1 Pro (on AI Pro)
• Code and analysis: 4 Argon
• Most complex work: 4 Carbon

Benchmarks and performance claims

All figures below are Google's claims.

• Artificial Analysis Intelligence Index: Google says Hydrogen surpasses Argon at its default high thinking level (51 vs. 50), reaching 54 at its highest xhigh thinking setting.
• Speed: Google says Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash at its high default setting. Partners saw it run 2x to 2.8x faster while reducing token usage by 12% to 23% compared to 3.8 Flash, according to Google.
• Terminal-Bench 4.1: 57%, versus 18.2% for 3.8 Flash.
• Reliability: Google says fabricated claims are cut by more than 5x, and that the model adapts cleanly when a tool call fails or a search returns empty.
• Agent Arena (arena.ai): ranks #5 on multi-step tasks such as deep research, data analysis, and broad codebase navigation.
• HLE-Diamond: 58.8%, which Google says leads its class.
• Vals Tax Agent Bench: 52.3%, ranked #1 overall.
• Vals Legal Research Bench: 55.8%, tied for #1.
• AutomationBench-AA: Google says it leads its class.
• Multimodal Knowledge Agent (MKA): 53.4%, which Google says outperforms other models on this new benchmark for agentic knowledge work over complex visual files.
• Agents' Last Exam v1.1: 71.6%, which Google says is ahead of other models on end-to-end professional workflows.

Google describes Hydrogen as a "major standalone upgrade" over Gemini 3.8 Flash, which launched on September 2. After that launch, Google focused on its frontier models, Argon and Carbon.

Gemini app, Search, and Antigravity

In the Gemini app, Google says Hydrogen follows tone, format, and rules more closely and produces clearer tables and visual breakdowns. It also more frequently calls the right tools (including Connected Apps) and skills. In Search (AI Mode), Google says Hydrogen improves intent understanding for multi-part queries, typos, and highly specific requests, with more visual responses and a wider range of links.

Google Antigravity gets an "Extra" thinking tier for Hydrogen and a new Orchestrator feature. Orchestrator switches dynamically between models such as Carbon and Hydrogen within one session. For parallel-agent workflows, Carbon can act as coordinator while Hydrogen powers worker and verifier agents across isolated workspaces. Google says this delivers Carbon-level quality with fewer tokens used.

What this means

Hydrogen is positioned as the everyday tier in Google's Gemini 4 lineup, sitting below Argon and Carbon in the app's model picker. Google's own Intelligence Index figures put it one point above Argon at the default setting, with the xhigh setting at 54, so the thinking level determines where it lands relative to Argon.

The Orchestrator design shows how Google intends Hydrogen to be used in agent workflows: Carbon coordinates and Hydrogen handles worker and verifier tasks. The stated aim is to reduce token spend. Google's speed, token-usage, and benchmark results come from the company and its partners, so developers should test them on their own workloads. Parameter count and training cutoff date have not been disclosed.