Google Ships Gemini 4 Hydrogen at $1.50 Per Million Input Tokens

Google launched Gemini 4 Hydrogen on November 12, 2026, priced at $1.50 per million input tokens and $7.50 per million output tokens — a model built for fast coding and agentic work rather than deep reasoning. The global rollout spans the Gemini API, AI Studio, Gemini for business, the Gemini app, Google Search, Antigravity and Gemini in Sheets for eligible subscriptions.

Hydrogen is adapted from Gemini 4 Carbon and slots into a clear division of labor. Carbon and Argon handle architectural judgment and extended reasoning; Hydrogen handles rapid execution — writing code, running tests, verifying changes, research and tool-using agents. It accepts text, images, audio and video, with a March 2026 knowledge cutoff that Google says may be less current in some domains.

What does the Orchestrator change for developers?

The new Orchestrator option in Google Antigravity operationalizes that split inside a single session. Carbon explores dependencies and produces a plan in a read-only phase; Hydrogen then edits code, updates packages and runs tests. If execution hits friction, control returns to Carbon to inspect state and repair the plan. Unlike a static router that commits an entire task to one model based on the opening prompt, Orchestrator can swap models as the work changes.

The Antigravity team reports near-Carbon task success on Terminal-Bench 4.1 while using 24% less quota and finishing 19% faster on average. For parallel work, Carbon coordinates while Hydrogen powers worker and verifier agents in isolated workspaces. Hydrogen is reaching paid Google AI Pro and Ultra subscribers in Antigravity now; Orchestrator starts with AI Ultra, with enterprise access to both through Gemini for business.

How fast is it, really?

The API documentation lists a rate-limited free tier and cached input at $0.075 per million tokens through implicit caching. Input, thinking and output share one 1,048,576-token window. Synchronous responses support up to 65,536 tokens per hop, with background execution or client-side continuation able to extend generation to the window's remaining capacity.

Google's reported numbers:

• Artificial Analysis Intelligence Index: 51 at high thinking versus Argon's 50, rising to 54 at xhigh
• Terminal-Bench 4.1: 57% versus 18.2% for Gemini 3.8 Flash
• Speed: 2.5 times faster than Gemini 3.8 Flash at high
• Fabricated claims: cut by more than fivefold

The model card carries a material qualification, though. Its cheating-adjusted Terminal-Bench 4.1 table lists Claude Sonnet 5.5 at 63.9%, ahead of Hydrogen's 57.0%. Hydrogen improves sharply on Google's previous Flash model without leading every competitor in that coding comparison.

Thinking levels and their trade-offs

Hydrogen offers four thinking levels: low, medium, high (the default) and a new xhigh. Google recommends high for most complex tasks and reserves xhigh for harder multistep problems where high falls short, noting that xhigh can overthink simpler prompts. The shared token budget makes these controls consequential — reasoning consumes capacity that could otherwise hold input or final output, and higher settings increase latency.

Safety and reliability picture

Google reports reductions of up to 81% in agentic overreach and 79% in under-service versus Gemini 3.8 Flash, plus greater resistance to indirect prompt injection. Those numbers matter when a model selects tools and acts across connected services — the Gemini app now draws on uploads and connected services including Gmail, Drive, Airtable and Dropbox, and Google says it follows tone and formatting instructions more closely and picks tools more reliably. Search gains better handling of multipart or typo-filled queries, more visual responses and expanded query fan-out.

The caveats are documented too. The model card's automated comparisons record a 5.57-percentage-point increase in unjustified refusals versus Gemini 3.8 Flash. Listed limitations include hallucinations, occasional slowness or timeouts, and potentially higher token use at elevated thinking levels. Free Gemini app users get Hydrogen only through automatic model selection when Google deems it appropriate.

The frontier-safety assessment found no material general-capability increase over Argon. Targeted tests placed Hydrogen below the biological-information and cybersecurity critical capability levels — Google's markers for severe-scale risk — while reaching alert thresholds in both domains. Fuller safety documentation is forthcoming.

The so-what: Hydrogen's real bet is that frontier planning belongs on hard stages while a cheaper, faster executor does the rest across mass-market products. Whether that arrangement preserves the reported efficiency gains in daily use — alongside the documented refusal and reliability limitations — is the question this rollout has to answer.