Google Launches Gemini 4 Hydrogen at $1.50/1M Tokens Across API, Search, and Antigravity Orchestrator

Google has launched Gemini 4 Hydrogen, the high-efficiency workhorse model in the Gemini 4 family designed to succeed Gemini 3.8 Flash across agentic coding, enterprise knowledge synthesis, and consumer search. Announced simultaneously by Google DeepMind SVP Koray Kavukcuoglu, Google Search VP Robby Stein, and the Google Developer Tools team—with day-one inference support on Vercel AI Gateway (google/gemini-4-hydrogen)—the model pairs a 1 million-token input window with up to 1 million output tokens at a flat rate of $1.50 per 1 million input tokens and $7.50 per 1 million output tokens.

Where Gemini 4 Carbon and Gemini 4 Argon target deep architectural reasoning and security analysis, Hydrogen is engineered as the fast, disciplined execution engine for daily repository iteration and multi-agent worker swarms. Despite operating in the Flash latency tier, Hydrogen surpasses Gemini 4 Argon on the Artificial Analysis Intelligence Index at its default high thinking setting (51 vs. 50) and climbs to 54 at its maximum xhigh reasoning tier.

Antigravity Orchestrator & Parallel Tool Batching for Developers

A centerpiece of the developer release is Orchestrator in Google Antigravity, an automated multi-model routing layer that dynamically switches between frontier planners and high-speed execution models within a single coding session. Instead of burning frontier token quota on routine file reads, test execution, and syntax verification, Orchestrator assigns architectural coordination to Gemini 4 Carbon while dispatching Gemini 4 Hydrogen worker and verifier agents across isolated workspaces in parallel.

According to Google DeepMind’s benchmark disclosures, that division of labor yields substantial production gains:

• Multi-Model Quota & Speed Gains: On Terminal-Bench 4.0, Antigravity Orchestrator matched near-Carbon task completion rates while consuming 24% less quota and finishing 19% faster on average.
• More Than 3x Terminal-Bench 4.1 Accuracy: Running standalone at its default high thinking level, Hydrogen finishes tasks 2.5x faster than Gemini 3.8 Flash while jumping from 18.2% to 57.0% on Terminal-Bench 4.1 and cutting fabricated claims in coding reports by more than 5x.
• Native Parallel Tool Batching: Rather than inspecting repository files sequentially, Hydrogen natively batches file-read and search tool calls in parallel. Across early enterprise testing on complex graph, SQL, and large codebase workflows, Hydrogen executed 2x to 2.8x faster while reducing total token usage by 12% to 23% compared to 3.8 Flash.
• Four Configurable Thinking Levels: Developers calling the Gemini API or Vercel AI SDK (via providerOptions.google) can tune reasoning depth across low, medium, high (default), and a new xhigh tier—paying only for the reasoning budget a specific step requires.

Gemini 4 Hydrogen vs. Gemini 3.8 Flash & Gemini 4 Family Comparison

To help engineering leads evaluate where Hydrogen fits alongside existing models and gateways, the table below synthesizes confirmed specifications, pricing, benchmark scores, and safety metrics across Google’s official model disclosures and Vercel’s AI Gateway changelog:

Capability / Metric	Gemini 4 Hydrogen	Gemini 3.8 Flash (Predecessor)	Antigravity Orchestrator (Carbon + Hydrogen)
API & Gateway Model ID	gemini-4-hydrogen / google/gemini-4-hydrogen	gemini-3.8-flash	Dynamic session routing in Google Antigravity
Context & Max Output Window	1M input tokens / up to 1M output tokens (flat rate)	1M input tokens / standard output decode	1M shared workspace context across coordinator & workers
API Pricing (Input / Output)	$1.50 / 1M input • $7.50 / 1M output (0% markup on Vercel)	Legacy Flash tier pricing	24% lower quota usage & 19% faster on Terminal-Bench 4.0
Configurable Thinking Levels	4 tiers: low, medium, high (default), xhigh	Standard thinking controls	Adaptive per subtask (Carbon planning + Hydrogen execution)
Artificial Analysis & Terminal-Bench	51 AA Index (54 at xhigh) • 57.0% Terminal-Bench 4.1	18.2% on Terminal-Bench 4.1 (2.5x slower task completion)	Near-Carbon frontier success rate on Terminal-Bench 4.0
Domain & Multimodal Benchmarks	58.8% HLE-Diamond • 52.3% Vals Tax (#1) • 53.4% MKA • 71.6% ALE v1.1	Baseline multimodal synthesis; 30%+ slower in Google Sheets canvas	Parallel worker synthesis across PDFs, code, and spreadsheets
Agent Alignment & Security	81% less overreach • 12x fewer misaligned tool actions • 2.3x Gray Swan IPI resilience	Baseline Flash guardrails	96.7% agent honesty rate • 99.9% CoT legibility (200K+ turns)

Global Google Search Upgrade & Workspace Multimodal Execution

Beyond developer APIs, Google Search VP Robby Stein confirmed that Gemini 4 Hydrogen is now live globally in Google Search for all users across every supported AI Mode language—marking what Google calls its largest single model upgrade to Search. Hydrogen’s stronger instruction following and tool calling upgrade Google Search’s simultaneous query fan-out engine: when a user submits a complex, multi-part natural-language question, Search dispatches parallel subtopic queries at once, surfacing a wider range of deep web links alongside dynamically generated comparison tables, visual breakdowns, and custom interactive tools.

In enterprise knowledge and productivity workflows, Hydrogen’s 1-million-token input and 1-million-token output capacity enables end-to-end document and spreadsheet synthesis without chunking truncations. It ranks #1 overall on the Vals Tax Agent Bench (52.3%), ties for #1 on the Vals Legal Research Bench (55.8%), and leads its class on the Multimodal Knowledge Agent (MKA) benchmark (53.4%) and Agents’ Last Exam v1.1 (71.6%). Inside Gemini in Google Sheets, Hydrogen generates interactive financial forecasting models more than 30% faster than prior models, while in the consumer Gemini app it connects across Gmail, Google Drive, Airtable, and Dropbox with fewer multi-turn corrections.

Agentic Safety, Prompt-Injection Resilience & Day-One Rollout

Because high-speed models frequently power autonomous background loops, Google DeepMind emphasized substantial alignment hardening in the Gemini 4 Hydrogen model card. Compared to Gemini 3.8 Flash, Hydrogen reduces agentic overreach by up to 81%, cuts under-service by up to 79%, and reduces misaligned actions or task shortcutting across coding and tool-use evaluations by 12x. Across more than 200,000 audited reasoning turns, the model maintained 99.9% chain-of-thought legibility with zero deceptive reasoning edits, achieved 2.3x higher resilience against indirect prompt injection on the Gray Swan IPI benchmark, and kept false medical/scientific refusals on HealthBench Pro under 0.4%.

Developers can call gemini-4-hydrogen immediately in Google AI Studio, Android Studio, and Vercel AI Gateway (using npx vercel ai-gateway setup to configure Claude Code, Codex, or Cursor with zero platform fee, including on Bring Your Own Key requests). Combined with yesterday’s Gemini Decisions API ($0.10/1M single-pass routing) and Gemini 4 Carbon, Google now offers a complete three-tier agentic stack: sub-250ms classification via Decision 1, fast $1.50/1M code and tool execution via Hydrogen, and deep frontier reasoning via Carbon.