Skip to content
ISSUE 001·LIVE·20:20 IL
← Journal/2026-11-13·9 min·google gemini

Gemini 4 Hydrogen: Google's fast execution model undercuts Argon on price and beats it on intelligence

Google launched Gemini 4 Hydrogen on November 12: a fast everyday coding and agentic model that scores 51 on the Artificial Analysis Intelligence Index at its default thinking level, above Gemini 4 Argon's 50, at 1.50 dollars per million input tokens and 7.50 per million output tokens. What the three-model Gemini 4 ladder means, how Orchestrator splits work in Antigravity, and what to verify.

By Harel Asaf·AI Builder·Tel Aviv

Google launched Gemini 4 Hydrogen on November 12, and it reshuffles the Gemini 4 lineup I wrote about two days ago. Gemini 4 Hydrogen is the fast execution model for everyday coding and agentic work, and per Google it scores 51 on the Artificial Analysis Intelligence Index at its default high thinking level, above Gemini 4 Argon's 50, while costing 1.50 dollars per million input tokens and 7.50 dollars per million output tokens, a quarter less than Argon on both sides. It is available on every tier, including free Gemini app users via auto model selection. The Gemini 4 family is now a three-model ladder: Hydrogen for speed, Argon for the default middle, Carbon for the hardest long-horizon work.

Where Gemini 4 Hydrogen sits in the lineup

Google's own framing per the launch post by Koray Kavukcuoglu, SVP at Google DeepMind: Carbon and Argon handle deep, long-horizon reasoning on the hardest problems, while Hydrogen is built for rapid execution across diverse agentic workflows. Practically, the ladder now looks like this:

  • Hydrogen: 1.50 / 7.50 dollars per million input / output tokens, flat across its full 1 million-token context, with up to 1 million output tokens. Everyday coding, debugging, iteration, multimodal knowledge work.
  • Argon: 2 / 10 dollars. Still positioned as the default for the majority of software engineering and knowledge work.
  • Carbon: 2 / 10 dollars introductory, doubling to 4 / 20 after November 30, 2026. The frontier model for multi-step coding, research, and cyber work via the Fairwind Program.

The notable tension: Hydrogen beats Argon on the cited intelligence index (51 vs 50, reaching 54 at its new xhigh thinking tier) while undercutting it on price. Unless Argon's real-world behavior differs from the index, Google has made its own default model hard to justify on paper.

Four thinking levels, one model

Hydrogen exposes four thinking levels: low, medium, high (default), and a new xhigh tier. Google says quality, latency and cost scale in a smooth progression, so you pay for reasoning depth only when a task needs it. For agent builders this is the more interesting control surface: a router that sends cheap steps to low and verification steps to xhigh is now a single-model design instead of a multi-provider one.

Orchestrator: multi-model sessions in Antigravity

Alongside Hydrogen, Google introduced Orchestrator in Google Antigravity: instead of one model per session, it switches dynamically between models like Carbon and Hydrogen. Google's numbers on Terminal-Bench 4.0: near-Carbon task success rates, 24% less quota consumed on average, 19% faster completion. For parallel work, Carbon can act as coordinator while Hydrogen powers worker and verifier agents across isolated workspaces. This is the same pattern xAI shipped this week in Grok Bot's teammate roles and that Anthropic pushes with subagents; Google is now baking it into the IDE instead of leaving it to your harness.

Performance claims, all vendor-reported

Against Gemini 3.8 Flash, Google claims Hydrogen finishes tasks 2.5 times faster at the default high setting, more than triples Terminal-Bench 4.1 scores (57% vs 18.2%), cuts fabricated claims by more than 5x, and recovers better from failed tool calls or empty searches. Partners testing complex graphs, SQL and data-heavy workflows saw 2x to 2.8x faster runs with 12% to 23% lower token usage versus 3.8 Flash. On public benchmarks Google cites: number 5 on Agent Arena, 58.8% on HLE-Diamond (leading its class), number 1 on Vals Tax Agent Bench (52.3%), tied first on Vals Legal Research Bench (55.8%), class lead on AutomationBench-AA, 53.4% on the new MKA multimodal benchmark, and 71.6% on Agents' Last Exam v1.1. All of these are Google's reports of its own and partner runs; independent reproductions do not exist yet.

Agent safety numbers worth reading

The safety section continues the Gemini 4 family's unusually specific reporting. Per Google: Hydrogen reduces agentic overreach by up to 81% and under-service by up to 79% versus 3.8 Flash, cuts misaligned actions and task shortcutting by 12x across tool-use and coding evals, is up to 2x less likely to bypass permission boundaries on impossible or blocked tasks, hits 96.7% on Google's internal agent honesty evaluation, preserves 99.9% chain-of-thought legibility with zero deceptive reasoning edits across more than 200,000 audited turns, and is more than 2.3x more resilient on the Gray Swan indirect prompt injection benchmark. Internal metrics, but the permission-boundary number is the one to track if you run agents with real tool access.

Availability

Per Google: rolling out globally now. Developers get it in the Gemini API via AI Studio and Android Studio, and in Antigravity for paid Google AI Pro and Ultra subscribers. Enterprises get it in Gemini for business via API or model dropdown, Antigravity, and Gemini in Google Sheets for eligible subscriptions. Consumers get it in the Gemini app model dropdown and in Google Search for all users; free app users get Hydrogen through auto model selection when it fits the prompt. The same announcement includes a separate Gemini Omni 1.2 Flash update for longer video clips, not covered here.

What I would check

My untested suggestions before switching defaults:

  • Run Hydrogen against Argon on your own task set before moving anything; a one-point index lead is inside noise, but the 25% price cut is real either way.
  • Test the thinking levels as routing knobs: measure where low and medium break on your workloads before wiring them into an agent loop.
  • If you use Antigravity, benchmark Orchestrator against your own Carbon-only and Hydrogen-only runs; the 24% quota and 19% speed numbers are Google's averages.
  • Recheck your Carbon evaluation from this week: if Hydrogen covers most of it at a quarter of the post-November-30 Carbon price, the calculus from two days ago already moved.
  • Treat the safety improvements as directional until the model card and external evals land; the Gray Swan comparison is versus 3.8 Flash, not versus Carbon's 0% claim.

What I would not conclude

Do not conclude Argon is dead; positioning and index scores are not workload truth. Do not assume the free-tier experience matches paid tiers; free users get Hydrogen only via auto selection. Do not treat Orchestrator's numbers as transferable to your own multi-model harness without measuring. And do not read 99.9% chain-of-thought legibility as a safety guarantee; it is an internal audit metric with an unpublished methodology.

Frequently asked questions

What is Gemini 4 Hydrogen?

Google's fast, efficient Gemini 4 model for everyday coding and agentic workflows, launched November 12, 2026. It sits below Argon and Carbon as the speed-optimized tier, with four thinking levels and a 1 million-token context window.

How much does Gemini 4 Hydrogen cost?

A flat 1.50 dollars per million input tokens and 7.50 dollars per million output tokens across its entire 1 million-token context. That is 25% cheaper than Gemini 4 Argon (2 and 10 dollars) and well below Carbon's post-introductory 4 and 20 dollars.

Is Gemini 4 Hydrogen better than Gemini 4 Argon?

On Google's cited Artificial Analysis Intelligence Index, yes by one point at default settings (51 vs 50), and 54 at the xhigh tier. Whether that beats Argon on your tasks needs your own eval; the price advantage, however, is unconditional.

What is Orchestrator in Google Antigravity?

A new mode that switches dynamically between models like Carbon and Hydrogen inside one session, with Carbon coordinating and Hydrogen running worker and verifier agents in isolated workspaces. Google reports near-Carbon success rates on Terminal-Bench 4.0 with 24% less quota and 19% faster completion.

What thinking levels does Gemini 4 Hydrogen support?

Four: low, medium, high (default), and a new xhigh tier. Google says quality, latency and cost scale predictably across them, so agent builders can route cheap steps and deep-verification steps to the same model at different levels.

Who can use Gemini 4 Hydrogen for free?

Per Google, free Gemini app users get Hydrogen through auto model selection when it fits the prompt, and it powers Google Search for all users globally. Direct model selection and API use require the paid surfaces.

Are the Gemini 4 Hydrogen benchmark and safety claims independent?

No. All figures cited here come from Google's launch posts: internal evaluations, partner testing, and its own runs of public benchmarks. Independent evaluations had not appeared at publication time.

Sources

Not read: the Gemini 4 Hydrogen model card, the Antigravity blog's Orchestrator details, and the Gemini API What's New guide. All performance and safety figures are Google's own; I did not test the model.

Related reading

  • Gemini 4 Carbon GA: what paying developers get today, and the November 30 price jump

    Google made Gemini 4 Carbon generally available on November 10: paid API access in AI Studio and Gemini for business, a /swarm multi-agent preview in Antigravity, Fairwind cybersecurity access, and introductory pricing that doubles after November 30, 2026. What changed since the October announcement, what Google claims, and what to verify yourself.

  • Qwen4-Max: Alibaba's 2.4T flagship API model claims 62% lower cost per task than its predecessor

    Alibaba's Qwen team released Qwen4-Max, the first model in the Qwen4 family: a 2.4-trillion-parameter sparse MoE with 71B active parameters, a 1M-token context window, thinking modes with tunable effort, and both OpenAI-compatible and Anthropic-compatible endpoints. Vendor-reported numbers show up to 49% fewer tokens on agentic coding tasks and a 62% drop in average cost per task versus Qwen3.8-Max. Input starts at 1.5 dollars per million tokens.

  • Inkling 2 Small: Thinking Machines ships a 594B open-weights MoE that runs on 17B active parameters

    Thinking Machines Lab released Inkling 2 Small with open weights under Apache 2.0: a natively multimodal 594B mixture-of-experts model with 17B active parameters, 1M tokens of context, and agentic-coding benchmarks that beat most open flash-class rivals. Self-hosting needs a GPU cluster; Ollama Cloud offers it at 0.30 dollars per million input tokens.

The build log

Follow the work.

New essays and prototypes, in your feed reader. Email subscriptions are paused while I connect a working delivery service. No email is collected here.

Follow via RSS