Gemini 4 Hydrogen Is Google's Cheapest Frontier Model — Here's How It Actually Compares

Google's own Facebook post for Gemini 4 Hydrogen says almost nothing: a blue graphic, the Gemini star, and the line "Intelligence & speed — for everyday coding and agentic work." No price, no context window, no benchmarks. Hydrogen launched on 12 November 2026 as the third Gemini 4 model in five weeks, and the comparison people actually want — Hydrogen against Argon and Carbon, and against Anthropic's Claude Fable 5.2 — is full of numbers that look comparable and are not. Every figure below was re-checked against the vendor's own announcement or pricing page, and anything that could not be re-verified was left out — the dropped numbers are listed at the end.

What Hydrogen actually is

Hydrogen is Google's execution model. In its launch post, Google DeepMind draws the line plainly: the frontier models Carbon and Argon "excel at deep, long-horizon reasoning on your hardest problems," while Hydrogen "is built for rapid execution across diverse agentic workflows" — writing code, running tests, verifying changes. It is both a standalone model for everyday work and the cheap execution engine inside multi-model workflows, where a frontier model plans and Hydrogen does the typing.

The facts, all from Google's announcement:

• Price: $1.50 per million input tokens and $7.50 per million output tokens, flat across its entire 1 million-token context window. The cheapest Gemini 4 by list price.
• Context: 1 million tokens of input, up to 1 million of output.
• Thinking levels: low, medium, high (default), and a new xhigh tier.
• Availability: rolling out globally from day one — the Gemini API and AI Studio, Android Studio's paid tier, Gemini for business, the Gemini app, Google Search for all users, and Antigravity for paid Pro and Ultra subscribers. It is the first Gemini 4 model free users meet.
• Model ID: gemini-4-hydrogen.

Google also announced Orchestrator in Antigravity, which switches between models inside one session — Carbon plans, Hydrogen executes. Google claims that on Terminal-Bench 4.0, Orchestrator reached near-Carbon success while consuming 24% less quota and finishing 19% faster. That is Google's own measurement, and it is the clearest statement of what Hydrogen is for: not the brain, the hands. Against its predecessor, Google claims Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash, more than triples Terminal-Bench 4.1 scores (57% vs. 18.2%), and cuts fabricated claims by more than 5x. On its own benchmark selection it ranks #5 on Agent Arena and leads its class on HLE-Diamond (58.8%). All Google's claims; none independently reproduced yet.

The Gemini 4 family, in one table

Table sources: Google's Hydrogen announcement (12 Nov 2026) for Hydrogen's price, context, thinking levels, availability and its Intelligence Index scores against Argon; Google's Carbon announcement and Google Cloud pricing docs for Carbon's $2/$10 intro and $4/$20 from 1 December; Google's Argon announcement and pricing page (as re-checked in launch coverage) for Argon's identical price schedule and unified window; Carbon's unified window from Google's Gemini API model documentation. Carbon's index figure (60) is the value in the Artificial Analysis index table that also lists Fable 5.2 at 65.

How to read the family table without fooling yourself

The $2/$10 price on Argon and Carbon expires on 30 November. From 1 December, both cost $4/$20. Hydrogen's $1.50/$7.50 carries no such expiry in anything Google has published — it is a flat rate, not an introductory one. After 1 December, Hydrogen is not 25% cheaper than Argon and Carbon, it is 62.5% cheaper on both input and output. If you are comparing list prices in a screenshot from this week, you are comparing a price that dies in two weeks against one that does not.

Per-token price is not per-task cost. Google says so itself about Carbon: it uses more thinking tokens than Argon on the same work, and thinking tokens are billed as output. The same applies to Hydrogen's thinking levels — xhigh buys the 54 by spending more tokens. A model at $1.50/$7.50 that thinks twice as long as a model at $2/$10 is not cheaper. The only honest comparison is cost per finished task on your own workload, and nobody has published one for Hydrogen yet.

The Intelligence Index is one battery, measured at different effort settings. Hydrogen's 51 is at its default high setting and 54 at xhigh; Fable 5.2's 65 is at Max effort. Same index, different settings, different dates — a rough ordering, not a precise ranking. Google's own benchmark charts are Google's selections, not independent measurements.

One caveat for the fine print: Google Cloud's model page lists Hydrogen as a short-term generally available release subject to a minimum 45-day retirement notice, with no tuning, no RAG Engine and no Live API support. That comes from the Cloud model page as reported in launch coverage — not seen first-hand on a Google page in this pass — but it matters for long-term builders: the cheapest tier is also the one Google can retire fastest.

Same week, different lanes: Decision 1 and H3-Image

Mazen also asked for Hydrogen against "the other two new ones" — the two launches the scout filed the same day, Gemini Decision 1 and MiniMax H3-Image. The honest answer is that they are not Hydrogen's rivals. They are different tools for different jobs, launched in the same week:

The comparison that matters here is architectural, not numerical: Decision 1 exists because agents spend most of their time choosing, not writing — it is the plumbing underneath a model like Hydrogen, not an alternative to it. H3-Image exists because image generation is a separate race with its own economics and licensing. Putting Hydrogen's $1.50 next to Decision 1's $0.10 or H3-Image's "50–70%" tells you nothing, because the tokens buy different things.

The quick Anthropic check

Anthropic's rival is Claude Fable 5.2, released 9 November 2026. This is the comparison where wrong numbers circulate most, so here it is, kept deliberately short and strictly sourced:

Read that table neutrally and it says something boring and true: Anthropic's frontier model costs roughly five times Hydrogen's list price per token and scores higher on the one independent index where both appear — 65 against Hydrogen's 51–54 — while Google's frontier models sit between them. A price–capability ladder, not a winner. The two numbers most often mangled in reposts: Fable 5.2's price is $8/$40, not $10/$50 (that was 5.1, before the 20% cut), and its FrontierCode 2.0 score of 49.1% is at Xhigh effort — at High effort it is 45.3%, essentially tied with Opus 5.5. Anyone quoting 49.1% without the effort level, or $10/$50 as the current price, is quoting stale numbers.

Who should pick which

• Everyday coding and agent execution, on a budget: Hydrogen. The cheapest Gemini 4, in front of free users, and the model Google itself frames as the worker. Start here; move up only when a task demonstrably needs more reasoning.
• The hardest, longest jobs — architecture, deep debugging, security, science: Carbon until 30 November, then re-price it. At $2/$10 it costs the same as Argon for a stronger model; at $4/$20 from December the token burn matters more than the rate.
• Most software engineering and knowledge work, per Google's own guidance: Argon — still Google's stated pick for the majority of demanding work.
• The frontier, on Anthropic's stack: Fable 5.2, at $8/$40 — five times Hydrogen's input price and more than five times its output price, for the highest independent index score of the four. Ask whether your task needs that top score, because the bill scales with it.
• Thousands of small judgements inside an agent: Decision 1 — a thousand questions for the price of a rounding error.
• Images: H3-Image, if you are outside its excluded territories and can live without inpainting and seeds for now.

How to read these numbers

• Introductory vs. flat: Argon and Carbon's $2/$10 ends 30 November 2026. Hydrogen's $1.50/$7.50 is flat. Fable 5.2's $8/$40 is a list price after a 20% cut. Three kinds of price, one table — check the date on the number.
• Input vs. output: every price here is two numbers. Output costs five times input on every Gemini 4 model, and thinking tokens are billed as output.
• Vendor benchmarks vs. independent ones: the only independent battery here is Artificial Analysis's Intelligence Index. Everything else — Terminal-Bench, FrontierCode, HLE-Diamond, Vals, CursorBench, AutomationBench — is a vendor's own selection, run on the vendor's harness.
• Effort settings change scores: Fable 5.2 scores 49.1% on FrontierCode 2.0 at Xhigh and 45.3% at High. Hydrogen scores 51 on the Intelligence Index at high and 54 at xhigh. A score without its effort setting is not a number, it is a rumour.
• Per-token vs. per-task: no independent cost-per-task figure exists yet for Hydrogen. Until one does, every "cheapest" claim here is about the rate card, not the bill.

Sources and classification

Classification: Official Announcement (Google, 12 November 2026) for Hydrogen's launch facts, with analysis. Launch facts and pricing are from Google's own announcement and pricing documentation; benchmark, latency and capability figures are the vendors' own claims, attributed as such above, with no independent verification published yet. Checked 13 November 2026.

• Google Blog, "Introducing Gemini 4 Hydrogen for Agentic Workflows" (Koray Kavukcuoglu, Google DeepMind, 12 November 2026) — read first-hand: Hydrogen's positioning vs Carbon and Argon; $1.50/$7.50 flat across the 1M-token context window; AA Intelligence Index 51 (high) vs Argon 50, 54 at xhigh; 1M input + up to 1M output; four thinking levels (low, medium, high default, xhigh); Orchestrator in Antigravity (near-Carbon success on Terminal-Bench 4.0, 24% less quota, 19% faster); vs 3.8 Flash (2.5× faster, Terminal-Bench 4.1 57% vs 18.2%, fabricated claims down >5×); Agent Arena #5, HLE-Diamond 58.8%, Vals Tax Agent Bench 52.3%, Vals Legal Research 55.8%, AutomationBench-AA, MKA 53.4%, Agents' Last Exam 71.6%; safety claims; availability across AI Studio, Gemini API, Android Studio paid tier, Gemini for business, Gemini app, Search for all users, Antigravity for Pro/Ultra: https://blog.google/intl/en-in/products/introducing-gemini-4-hydrogen/
• Google, Facebook post by Google (@Google), 12 November 2026 — the promo graphic Mazen sent: "Gemini 4 Hydrogen — Intelligence & speed — For everyday coding and agentic work": https://www.facebook.com/Google/posts/pfbid09kiNU5kDBUUHAL3PD9BfVMuYLtU6RZTyu4MZVMhqCVX2Gx7DNLxT2iELpYpw5ejAl
• Google Blog, "Gemini 4 Carbon is now available" (Koray Kavukcuoglu, 10 November 2026) — read first-hand for the Carbon piece: $2/$10 introductory, doubling to $4/$20 after 30 November 2026; Argon as the pick for most work, Carbon for the hardest jobs, Carbon burns more tokens: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/
• Google Cloud, Gemini pricing documentation — Gemini 4 schedule confirming $2/$10 introductory through 30 November 2026 and $4/$20 from 1 December 2026 for the Gemini 4 frontier line: https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing
• Google Gemini API model documentation (as re-checked in launch coverage, 11–12 November 2026) — Argon and Carbon: 1,048,576-token unified context window, 65,536-token default synchronous output, up to 1M via long decode; Carbon thinking levels low/medium (default)/high; model IDs gemini-4-argon, gemini-4-carbon.
• Anthropic, "Introducing Claude Fable 5.2 and Claude Mythos 5.2" (9 November 2026) — read first-hand for the Fable piece: $8/$40 (20% cut from $10/$50); up to 30% faster output; fast mode $16/$80, up to 2.5×; FrontierCode 2.0 49.1% (Xhigh) vs Fable 5.1 32.4%, Opus 5.5 45.5%, GPT-6 Astra 42.8%, Gemini 4 Carbon 41.1%: https://www.anthropic.com/claude-fable-and-mythos-5-2
• Artificial Analysis, Intelligence Index v5 — Fable 5.2 ranked #1 with a score of 65 (Max effort); Carbon 60 and Opus 5.5 59 in the same index table (corroborated by The Decoder's coverage of the index).
• Google Blog, "Gemini Decisions API: Fast, calibrated multimodal decisions" (Min Choi and Felix Yu, Google DeepMind, 11 November 2026) — read first-hand for the Decision 1 piece: $0.10 per million input tokens, $0 output; up to 1,000 questions in one pass; 256K-token shared input; 150–250 ms latency: https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/
• MiniMax Blog, "MiniMax H3-Image: Bringing 4K Image Generation into the 10-Second Era" (10 November 2026, 16:23 UTC) — read first-hand for the H3-Image piece: 33.1B parameters; 4096×4096 in under 10 seconds; up to 16 reference images; pricing 50–70% of comparable mainstream offerings; 50% launch-week discount to 17 November, 16:00 UTC: https://www.minimax.io/blog/minimax-h3-image
• MiniMax H3-Image model card and GitHub repository — 68.65 GB BF16 checkpoint; Community License Agreement excluding the US, EU, UK and South Korea; no masked inpainting, transparent backgrounds, user-specified seeds or preview images: https://huggingface.co/MiniMaxAI/MiniMax-H3-Image
• 3BQRIINO live site, sitemap checked 13 November 2026 — internal links verified live before linking: Carbon (/blog/2026-11-11-gemini-4-carbon/), Decision 1 (/blog/2026-11-12-gemini-decision-1/), H3-Image (/blog/2026-11-12-minimax-h3-image/), Fable 5.2 (/blog/2026-11-10-claude-fable-mythos-5-2/).