Google Launches Gemini 4 Hydrogen at $1.50 and $7.50 per Million Tokens Across Search, API, and Antigravity

Google DeepMind on November 12, 2026, positioned Gemini 4 Hydrogen as the fast, efficient face of its fourth-generation model family—a deliberate contrast to the deeper reasoning tiers Carbon and Argon. Hydrogen is priced at a flat one dollar and fifty cents per million input tokens and seven dollars and fifty cents per million output tokens across the full one-million-token context window, with cached inputs at seven and a half cents per million. Google said the model would appear simultaneously in Google AI Studio, the Gemini API, the consumer Gemini app, Google Search, business tiers, and the Antigravity agent environment.

Positioning in the Gemini 4 lineup

Koray Kavukcuoglu, senior vice president at Google DeepMind and chief AI architect, framed Hydrogen as the workhorse for agentic coding and multi-step automation. Carbon and Argon remain the choice for long-horizon reasoning, but Hydrogen targets throughput-sensitive tasks: parallel tool calls, repository refactors, document batching, and customer-service agents that must answer within seconds.

On the Artificial Analysis Intelligence Index, Google said Hydrogen at its default high thinking level scores fifty-one against Argon’s fifty, and fifty-four at the new xhigh setting. That benchmark marketing matters because many enterprises still choose models by leaderboard rank even when latency and price dominate real deployments.

Pricing mechanics developers should model

Unlike earlier Gemini releases that discounted short prompts, Hydrogen uses flat per-million-token pricing regardless of context length. Output billing includes thinking tokens, an important detail for teams that enable higher reasoning modes by default.

Batch and Flex tiers halve both input and output rates to seventy-five cents and three dollars seventy-five cents per million tokens, respectively. Priority tiers cost more for guaranteed capacity. Context caching offers a ninety-five percent discount on repeated prefixes, but only through implicit caching—explicit cache APIs are not supported on Hydrogen.

Grounding with Google Search remains priced per query after a monthly free allotment, and developers must budget for multiple search calls per user turn when agents research autonomously.

API and thinking-level differences

Hydrogen introduces a fourth thinking level, xhigh, alongside low, medium, and high. Carbon and Argon default to medium thinking and cap at high. Teams migrating workloads must retest latency and cost because default reasoning depth changed.

The model also supports sixty-five thousand five hundred thirty-six tokens per non-streaming hop, double the thirty-two thousand seven hundred sixty-eight limit on Carbon and Argon REST calls. Long-form generation without chunking reduces glue code but increases blast radius when a run fails midway.

Google documentation encourages hybrid routing: Carbon for planning turns, Hydrogen for execution turns within a single session. That pattern mirrors how human teams separate architecture reviews from implementation sprints.

Search, Antigravity, and consumer surfaces

Shipping Hydrogen into Google Search on day zero signals Google’s intent to make its fastest model the default for interactive answers, not only developer APIs. Search integration implies tighter coupling between retrieval, ranking, and generation—areas where Google’s data moat still differs from pure model vendors.

Antigravity, Google’s agent orchestration layer, gains a model tier aligned with paid Google AI Pro and Ultra subscribers. The bundling strategy pressures standalone agent frameworks to justify premiums when platform owners include competitive models in existing subscriptions.

For enterprises, the practical question is whether Search-grounded Hydrogen answers reduce the need for custom RAG stacks, or whether compliance teams will block cloud models from touching proprietary indexes regardless of convenience.

Competitive context

VentureBeat’s launch-day comparison table placed Hydrogen at nine dollars combined per million tokens (input plus output at typical ratios), squarely in the mid-market against Claude Sonnet-class models and above many Chinese flash offerings, but below frontier Opus- and Astra-tier pricing.

LiteLLM and other gateways added day-zero routing, meaning startups can trial Hydrogen without rewriting client SDKs. That accelerates bake-offs during the same week Anthropic markets its IPO and OpenAI navigates Capitol Hill on export controls.

Google’s pricing also sunsets introductory rates on Gemini 4 Argon and Carbon at the end of November, after which those models jump to four dollars and twenty dollars per million tokens. Hydrogen becomes the economical on-ramp into Gemini 4 for teams that do not need maximum reasoning depth on every call.

Takeaways for engineering leaders

Treat Hydrogen as a default for high-volume agent steps, not as a replacement for every Carbon workflow. Re-benchmark retrieval-heavy pipelines because Search grounding fees can dominate token costs.

Update observability to tag thinking levels and cache hits; without those dimensions, FinOps teams will misattribute spikes to user growth rather than configuration changes.

Finally, plan for policy review: Google’s contributor tier still trades discounted inference for training data rights. Regulated industries may need enterprise contracts that explicitly opt out, even when using Hydrogen through a paid API key.

Enterprise procurement checklist

Procurement teams should request written confirmation of data retention policies for Hydrogen calls routed through Search grounding, because user queries may trigger multiple paid search fetches. Legal teams should compare Google’s standard cloud terms with sector-specific rules in finance and healthcare before enabling Antigravity agents on customer data.

FinOps leads ought to model three scenarios: all-high thinking, mixed Carbon-plus-Hydrogen routing, and Flex-only batch processing. Each produces different spend curves; without scenario planning, quarterly true-ups surprise budget owners.

Security architects must verify whether implicit caching could leak prompt prefixes across tenants in shared projects. Google positions caching as a customer benefit, but multi-team organizations need project boundaries and key rotation policies documented before production cutover.

Migration notes from Gemini 3.8 Flash

Teams on promotional Gemini 3.8 Flash pricing—which remains at seventy-five cents and three dollars seventy-five cents per million tokens through December 31, 2026—face a January rate hike identical to Hydrogen’s list price. Google is effectively training customers on Hydrogen APIs now so January invoices do not force emergency refactors.

SDK changes are modest: model string updates to gemini-4-hydrogen, explicit thinking_level parameters, and optional Interactions API paths for multi-turn agents. Regression suites should include tool-calling transcripts because Hydrogen’s higher non-streaming hop limit changes failure modes for oversized JSON payloads.

Why Search integration matters for rivals

When Hydrogen powers interactive answers in Search, third-party publishers see traffic implications within days, not quarters. Marketing analytics teams will need to separate AI-referral segments from classic organic clicks. Competitive intelligence units at Microsoft and OpenAI will watch whether Hydrogen citations favor Google-owned properties or maintain neutral sourcing—an unresolved policy question every regulator has asked in slightly different words since 2023.

Closing view

Hydrogen is not a quiet mid-cycle refresh. It is Google’s attempt to align consumer Search, developer APIs, and paid agent platforms on one economical inference tier before rival models reset price expectations again. Teams that treat the launch as a simple model-string swap will misforecast spend; teams that rehearse routing, grounding fees, and caching rules this week will enter 2027 with cleaner unit economics than competitors still guessing from blog posts alone.

Reliability and quota planning

Platform teams should model Hydrogen rate limits alongside price. Free-tier throttling can masquerade as model quality issues when internal demos hit shared API keys. Separate projects per environment, alert on 429 responses, and document fallback to Flex mode for non-interactive jobs. Operations runbooks written for Gemini 3.x must note the doubled non-streaming hop limit so on-call engineers do not split responses unnecessarily.