Gemini 4 Hydrogen Ships at $1.50/$7.50 and Outranks Argon

The short version

On November 12, 2026, Google announced Gemini 4 Hydrogen and shipped it the same day — no preview, no waitlist. The name suggests "Gemini 4 Flash," which is the wrong read. Hydrogen is the execution model that sits above Argon on aggregate intelligence while costing a fraction of Carbon.

Google's own numbers:

51 on the Artificial Analysis Intelligence Index at the default high thinking level — versus 50 for Argon, the mid-tier reasoning model. It reaches 54 at the new xhigh tier.
57% on Terminal-Bench 4.1, against 18.2% for Gemini 3.8 Flash.
2.5x faster task completion than 3.8 Flash at the default setting, with fabricated claims cut by more than 5x.
$1.50 per million input tokens / $7.50 per million output tokens, flat across the full 1M-token context window.

Google also quietly retired the Flash brand. Hydrogen is not the cheap tier. It is the default.

The pricing story nobody leads with

Comparing $1.50/$7.50 to 3.8 Flash's current $0.75/$3.75 makes Hydrogen look twice as expensive. That comparison expires in seven weeks:

Gemini 3.8 Flash moves to $1.50/$7.50 on January 1, 2027.
Gemini 4 Carbon and Argon move to $4/$20 on December 1, 2026 (up from $2/$10).
Gemini 4 Hydrogen stays at $1.50/$7.50 with no announced change.

So from December 1, Hydrogen's output price is one quarter of Carbon's. Cached input is $0.075 per million — a 95% discount, implicit caching only. Batch and Flex tiers are half price at $0.75/$3.75; Priority is $2.70/$13.50.

If you have routine work running on Carbon today, that is a 4x cost reduction sitting there. Not because Hydrogen is weaker on the task, but because Carbon is about to double.

Orchestrator: dynamic routing instead of a static router

The more interesting release is Orchestrator in Google Antigravity, which pairs Carbon and Hydrogen inside a single session.

The usual industry approach is a static model router: guess task difficulty from the first prompt, route the whole task to the cheapest model that can plausibly finish it. That breaks down as agents run longer, because the right model changes mid-trajectory — and you only discover that after the task evolves.

Orchestrator does bidirectional handoff instead:

Carbon plans. A read-only phase to explore dependencies, resolve ambiguities, and build a concrete plan with explicit re-planning triggers.

Hydrogen executes. Write edits, update packages, run tests.

Hydrogen hands back on friction. Carbon inspects the state, repairs the plan, resumes execution.

On Terminal-Bench 4.1, Google reports near-Carbon task success with 24% less quota consumed and 19% faster completion. For parallel agent work, Carbon coordinates while Hydrogen powers worker and verifier agents across isolated workspaces.

Availability is plan-dependent, and the details matter:

Google AI Ultra: both Orchestrator (Carbon) and Orchestrator (Argon).
Paid Google AI Pro: Orchestrator (Argon) only.
Enterprise: both, via Gemini for business.

Hydrogen itself is available to paid AI Pro and Ultra subscribers in Antigravity, with four thinking levels: low, medium, high (default), and the new xhigh. Antigravity exposes an additional "Extra" tier.

What the benchmark claims miss

Rankings matter less than the behavioral descriptions, because that is where your daily loop actually changes:

Parallel file reads. Hydrogen natively batches file-read and search calls in parallel rather than inspecting files one at a time. Across complex graphs, SQL, and data-heavy enterprise workflows, Google says partners observed 2x to 2.8x faster runs with 12% to 23% lower token usage versus 3.8 Flash.
Surgical edits. It fixes the diff that matters without rewriting unrelated files, and outperforms same-tier models on SWE-sweep where you have to find and repair bugs without breaking working code.
Graceful failure. It adapts when a tool call fails or a search comes back empty.

External benchmarks are in the announcement too: #5 on Agent Arena, HLE-Diamond 58.8%, #1 on Vals Tax Agent Bench (52.3%), tied for #1 on Vals Legal Research Bench (55.8%), 53.4% on the new Multimodal Knowledge Agent benchmark, and 71.6% on Agents' Last Exam v1.1. With 1M input and up to 1M output tokens, synthesizing dense PDFs, spreadsheets, audio, and recorded video into a cited report is now a single call.

Read the safety numbers, but read them as vendor numbers

For agentic deployments, "does it do things I did not ask for" matters more than accuracy. Google reports, relative to 3.8 Flash:

Overreach in agentic tasks reduced by up to 81%; under-service by up to 79%.
Misaligned actions and task shortcutting across tool-use and coding evals cut by 12x.
Up to 2x less likely to bypass permission boundaries on impossible or blocked tasks; 96.7% honesty rate on internal agent evaluation.
99.9% chain-of-thought legibility with zero deceptive reasoning edits across 200,000+ audited turns.
2.3x+ more resilient on the Gray Swan indirect prompt injection benchmark.
False refusals on HealthBench Pro under 0.4%.

Every one of those is a Google-measured figure on Google's own internal harness. The Agent Arena and Vals results are third-party, but the honesty and alignment numbers are not. Treat them as directional.

The Search rollout is the part your business will notice

Hydrogen is now the default in the Gemini app and is rolling out to Google Search for all users globally. Google calls it the largest model upgrade to Search to date, with three specific changes:

Better intent parsing for multi-part queries and heavy typos.
More visual, scannable responses.
An upgraded query fan-out technique that surfaces a broader range of links.

That third point matters if you publish anything. As link selection moves into the model, the practical question shifts from "how do I rank" to "can this model parse and cite my page." Pages that render poorly to a parser get skipped regardless of quality.

Three things to do this week

Move routine work off Carbon before December 1. $2/$10 becomes $4/$20; Hydrogen stays at $1.50/$7.50. For repetitive long-prompt workloads, add caching at $0.075 per million input.

Run Orchestrator for a week and log the handoffs. The 24%-less-quota figure is Google's, on Google's harness. Measure it on one of your repos. On Pro you only get Orchestrator (Argon), which is a different pairing than the headline.

Do not default to xhigh. The gap between high (51) and xhigh (54) is three index points purchased with latency and tokens. Run high, escalate per-task on failure.

Bottom line

Hydrogen is not a discount model. It is the model that determines what Gemini 4 costs you from December onward, and it outscores the mid-tier reasoning model while undercutting Carbon's post-increase output price by 4x. Dropping the Flash name was the honest signal: this is the standard tier now, not the light one.

The split to build around is simple. Carbon decides, Hydrogen executes. Teams that structure their pipelines that way before December 1 pay the least and wait the shortest.