Strategy

Gemini 4 Hydrogen Prices a Flagship Flat Across 1M Context — and Confirms the Agent Bill Is Set by the Tier

By IdeaBosqueLast updated: November 12, 2026

Key takeaways

Gemini 4 Hydrogen (Nov 12, 2026) is priced flat at $1.50/$7.50 per million tokens across the entire 1M-token context window — no long-context premium, and no announced intro window, arriving two days after Carbon's GA at an intro $2/$10 that expires November 30 (Google).
Artificial Analysis's 95% cached-input discount blends Hydrogen to $1.10/1M at a 7:2:1 ratio — cached input bills at $0.075/1M, which is the rate that matters for always-on agents re-reading the same context every loop (Artificial Analysis).
The AA cost per Intelligence Index task is $3.16 (xhigh) — 4x GPT-6.1 Sol's $0.74 and 40% of Carbon's $7.76 — because the xhigh tier generated 304M output tokens across the Index runs, ~4x the 72.8M median; list price is not the bill, output volume is (Artificial Analysis).
Orchestrator in Google Antigravity routes Carbon and Hydrogen inside one session and hit near-Carbon task success on Terminal-Bench 4.0 while consuming 24% less quota and finishing 19% faster — model routing, the discipline behind every cost-effective production agent, is now a boxed platform feature (Google).
Hydrogen ships with no dated pricing cliff; Carbon's is November 30 — a two-week $2/$10 window ahead of a doubling to $4/$20. Contract at post-intro rates and put every discount expiry on the procurement calendar, regardless of which vendor's rate card you hold.

On November 12, 2026, Google introduced Gemini 4 Hydrogen: a flagship-class reasoning model framed explicitly "for Agentic Workflows," priced at $1.50 per million input tokens and $7.50 per million output tokens — flat across the entire 1M-token context window, with no long-context premium tier. It arrived 48 hours after Carbon's GA on November 10 at an introductory $2/$10 that expires November 30, and two days before Carbon's rate doubles to $4/$20 on December 1. Google shipped two flagship-family API models in three days, priced the newer one below the older one's intro, and framed the newer one as the execution tier for everyday agent work.

Three things make the day-0 rate card worth reading closely rather than skimming: the flat pricing removes the long-context premium that inflates agent input bills; the bundled Orchestrator in Antigravity turns model routing — until now a discipline each team assembled itself — into a platform feature; and the independent cost data shows the third-party bill is set by a variable none of the headline prices capture. This article walks the pricing mechanics, the cost-per-task evidence, what Orchestrator does and does not cover, and the procurement calendar the family now sits on.

What shipped on day 0

Google positions Hydrogen against its own flagships: Carbon and Argon "excel at deep, long-horizon reasoning on your hardest problems," while Hydrogen is "built for rapid execution across diverse agentic workflows" (Google). Measured intelligence places it precisely there: it scores 51 on the Artificial Analysis Intelligence Index at its default high thinking level versus Argon's 50, and 54 at the new xhigh setting (Artificial Analysis) — the level where Anthropic's Sonnet 5.5 and OpenAI's GPT-6.1 Terra cluster, not the 60-point tier where Carbon (high) sits.

Four thinking levels ship with the model — low, medium, high (default), and xhigh — on the stated principle that "you only wait and pay for the reasoning depth a task requires." Against the model it replaces in the everyday tier, Gemini 3.8 Flash, Google reports Hydrogen finishes tasks 2.5x faster at the high default while more than tripling Terminal-Bench 4.1 scores (57% versus 18.2%) and cutting fabricated claims by more than 5x (Google). Rollout covered all surfaces on day 0: AI Studio and the Gemini API, Gemini for business, the Gemini app, Google Search, and Antigravity for paid AI Pro/Ultra subscribers. Native modality limits are 1M input tokens and up to 1M output tokens.

What flat-across-1M pricing changes

The rate card's most consequential property is the one that reads as a footnote: flat pricing across the context window. Most frontier models bill higher rates as context grows; a hybrid always-on agent — an RFQ or monitoring loop carrying a 300K-token working set of catalogs, contracts, and prior turns — inherits the premium on every call. A flat $1.50 input rate across the full window removes that tier boundary: the same bill at 200K tokens as at 20K.

The cache layer is where the execution-loop economics actually settle. Artificial Analysis lists Hydrogen with a 95% cache discount — cached input at $0.075/1M — and its blended rate at a 7:2:1 cache/input/output ratio works out to $1.10 per million tokens (7×$0.075 + 2×$1.50 + 1×$7.50, divided by 10). An agent loop that re-reads the same long context on every turn sits close to that blend, which is the shape of a monitoring or reconciliation agent that was economically prohibitive two years ago and is now rounding error at per-task scale.

The honesty marker is what the pricing announcement does not contain: an expiry date. Carbon's November 10 GA shipped with a two-week intro window and a published doubling. Hydrogen's announcement states the rate without qualification — but as this site tracked with Carbon's dated cliff, intro-adjacent rates that expire on a published date are the family's operating pattern, and the absence of a cliff is a fact about today's announcement, not a commitment about December's.

List price is not the bill: verbosity sets the task cost

The independent cost measurement is the part of the day-0 data buyers should weight most, because it inverts the rate card's promise. Artificial Analysis costs Hydrogen (xhigh) at $3.16 per Intelligence Index task — more than four times GPT-6.1 Sol's $0.74, despite Hydrogen's list rates being comparable to or below the models around it. The cause is in the verbosity row: across the Index runs, the xhigh tier generated 304M output tokens — roughly four times the 72.8M median for its class. At $7.50/1M output, thinking tokens dominate the bill. AA's verdict: "very verbose," ranked 76th of 77 on token use.

Put the family side by side on the same per-task basis (Artificial Analysis, Artificial Analysis, and this site's earlier per-task exhibit set):

Model (tier)	List $/1M in–out	AA cost per Index task	AA Intelligence Index
GPT-6.1 Sol (max)	—	$0.74	—
Gemini 4 Argon (promo $2/$10)	$4/$20 discounted 50%	$1.99	50 (high)
Gemini 4 Hydrogen (xhigh)	$1.50/$7.50 flat	$3.16	54
Gemini 4 Argon (post-promotion)	$4/$20	$3.98	50 (high)
Gemini 4 Carbon (high)	$4/$20 after Nov 30	$7.76	60

The same 50s-class intelligence bills at $0.74, $3.16, or $7.76 per task depending on which vendor's verbosity and pricing strategy carries it. Hydrogen's $3.16 is a respectable number — it undercuts Carbon's post-cliff per-task cost by ~59% for six points of index score — but it is not the bargain the $1.50/$7.50 headline suggests, and the gap between the two is a property of output volume, not of any rate a buyer can select directly. The procurement rule stays: contract on cost per task measured on your own workload mix, not on the rate card. The Sonnet 5.5 token-efficiency exhibit showed the same inversion from the token-count direction; Hydrogen shows it from the thinking-tier direction.

Three days of flagship releases, one rate ladder — the chart below puts the list prices beside what an Intelligence Index task actually bills:

Orchestrator: model routing becomes a bundled platform feature

The release's most interesting architecture row is not a model at all. Orchestrator in Google Antigravity dynamically switches between Carbon and Hydrogen within a single session — deep-reasoning steps to Carbon, execution and verification steps to Hydrogen. On Terminal-Bench 4.0, Google reports Orchestrator reached near-Carbon task success while consuming 24% less quota and completing tasks 19% faster, with Carbon able to coordinate worker and verifier agents powered by Hydrogen across parallel workspaces (Google).

That is the inference-economics thesis — the cheapest capable model per task is a configuration decision — shipped as a vendor box rather than assembled from routing code. Its reach is the honest boundary: Orchestrator governs traffic inside Antigravity's surfaces. A team calling the Gemini API directly, or building on Bedrock, Vertex, or a model gateway, still owns the routing decision — which model handles which step, what the fallback does when a tier fails, and how the substitution risk is controlled. The DeepSeek V4.1-Flash auto-routing episode is the failure case for implicit routing: when a provider silently substitutes a cheaper model, or an orchestrator switches tiers, the buyer's cost model and output quality both move without notice. Bundled routing narrows that risk only inside the bundle; it converts it into a vendor-trust question rather than an architecture question.

The December procurement calendar

Three rate cards now coexist in one vendor family, and only one of them is currently honest about its future:

Carbon: $2/$10 intro, expires November 30, then $4/$20. The cliff is published, dated, and doubling. An agent fleet sized on intro rates silently reprices at month end unless the cost model is scheduled against the date.
Argon: $4/$20 rate card, currently discounted 50% to $2/$10. The discount structure is the same repricing risk in a different form — a discount without a published end date still ends.
Hydrogen: $1.50/$7.50 flat, no announced expiry. The cleanest rate card of the three today. Its cache economics ($0.075/1M cached input) are the strongest fit in the family for always-on loops.

The pattern to internalize is not the numbers; it is the cadence. Google moved the flagship family's price structure twice inside three days. Any commitment priced on today's rate card is a commitment priced on a moving target — so the cost model should run at post-intro rates by default, carry every discount expiry as a dated calendar entry, and treat an unchanged rate three weeks from now as evidence, not promise.

Vendor-reported safety numbers, labeled as such

Hydrogen's announcement carries a dense safety block worth cataloging for later verification, with the standing caveat that every figure is Google-reported and not independently audited: agent overreach reduced by up to 81% and under-service by up to 79%, misaligned actions and task shortcutting cut 12x across tool-use and coding evaluations, up to 2x less likely to bypass permission boundaries on impossible or blocked tasks, a 96.7% honesty rate on Google's internal agent honesty evaluation, 99.9% chain-of-thought legibility with zero deceptive reasoning edits across more than 200,000 audited turns, and 2.3x more resilience on the Gray Swan IPI benchmark against indirect prompt injection (Google).

For a buyer running an execution-tier model inside a governed agent, these are diligence inputs with a verification plan attached: the numbers name the failure modes that matter (overreach, shortcutting, prompt injection, deceptive reasoning) and give measurable starting points for the questions the governance checklist already asks. If 96.7% is the vendor's honesty rate, the buyer's question is what the remaining 3.3% did, under whose monitoring, and whether the audited-turn methodology is published.

What this changes for a buyer's cost model

Long-context input bills drop structurally. A flat $1.50 input rate across 1M tokens, with a 95% cached tier beneath it, removes the context-premium line item from always-on agent cost models — until the next undisclosed repricing.
The execution tier has a real incumbent now. Hydrogen at 54 on the AA Index at $3.16/task is the cheapest self-consistent execution tier Google has shipped at flagstaff capability; compare candidates per task, not per token.
Routing is a procurement question twice over. Once for whatever you assemble yourself (which step, which model, which fallback), and once for what the vendor's own Orchestrator decides inside its surfaces — both need the substitution-risk controls.
Re-verify the family's rate cards before any December commitment. Carbon's doubling lands November 30–December 1; Hydrogen's rate card is three days old.

The bottom line: Hydrogen is the strongest execution-tier rate card in the Gemini 4 family and the first flagship priced flat across 1M context — but Artificial Analysis's $3.16 per task against Sol's $0.74 is the day-0 reminder that the model you pay for is the one that thinks out loud in output tokens. Price the workload, not the rate card.

The pattern generalizes past frontier-model shopping. The same structure — an execution tier priced for loop volume, a reasoning tier priced for judgment, routing deciding which bill each step lands on, and per-task cost as the unit of contract — is exactly the shape of a production B2B agent's model layer, whether the tiers are Google's or assembled across labs on a gateway.