Gemini 4 Carbon vs Argon vs Hydrogen: Google’s Gemini 4 Models Compared

By stackcone · November 2026

Google now has three Gemini 4 models, and their names (Carbon, Argon, Hydrogen) say nothing about which is which. Here is the short version: Carbon thinks hardest, Argon is the long-task middle, and Hydrogen is the fast, cheap one that replaces Gemini 3.8 Flash.

All three share the same architecture and a 1,048,576-token context window. They differ mostly in how much thinking they spend per turn, which drives quality, speed and price. And two of them are about to double in price, so the timing matters.

Hydrogen price / 1M
$1.50 / $7.50
Carbon & Argon now
$2 / $10
Carbon & Argon from Dec 1
$4 / $20
Context window (all)
1,048,576

The three models at a glance

Gemini 4 CarbonGemini 4 ArgonGemini 4 Hydrogen
RoleDeepest reasoning, securityLong-horizon tasks, fewer tokens than CarbonFast everyday model, Flash successor
ReleasedOct 28 (cyber defenders); GA Nov 10Sept 30 (limited); now on paid APIGA Nov 12, 2026
Price per 1M (in / out)$2 / $10 intro$2 / $10 intro$1.50 / $7.50
From Dec 1, 2026$4 / $20$4 / $20Unchanged
Cached inputImplicit onlyImplicit only$0.075 per 1M (implicit)
Batch / Flex——Half price
Free accessNoGoogle AI Pro appFree API tier; free Gemini app (auto selection)
Image generation / Live APIYesYesNo
Model IDgemini-4-carbongemini-4-argongemini-4-hydrogen

Independent scores

Artificial Analysis Intelligence Index. Carbon and Argon at max effort (index v5.0.1); Hydrogen and 3.8 Flash per Google/VentureBeat.

The surprise is how close Hydrogen gets: at high reasoning it scores one point above Argon, and at xhigh it clears Argon by four, for less money. On LMArena’s blind-vote leaderboard, though, Carbon (1526) and Argon (1522) held the top two spots on November 9. People like how the bigger models write.

BenchmarkCarbonArgonHydrogenNote
AA Intelligence Index605051 (high) / 54 (xhigh)Independent
LMArena (Nov 9)1526 (#1)1522 (#2)—Blind user votes, preliminary
AutomationBench53.9%51.3% (at launch)—Anthropic and Google tables
Terminal-Bench—57.4% (TB 4.0)57% (TB 4.1)Google’s evaluations
ARC-AGI-2—91.3%—November roundup
BenchLM coding rank#2—#6Hydrogen also #6 agentic

When to use each one

Gemini 4 Hydrogen: the default

Start here. Hydrogen handles everyday coding, agents and multimodal documents, completes tasks 2.5x faster than Gemini 3.8 Flash, and scored 57% on Terminal-Bench 4.1 against Flash’s 18.2%. Google also claims more than 5x fewer fabricated claims than 3.8 Flash. It beats Claude Haiku 5.5 comfortably (AA Index 51 vs 36) but trails Claude Sonnet 5.5 on Terminal-Bench (57% vs 63.9% in Google’s table). See our mid-tier comparison.

Gemini 4 Carbon: the hard problems

Use Carbon for hard architecture questions, subtle bugs, scientific reasoning and security audits. It was built first for finding, validating and patching vulnerabilities, and it leads Anthropic’s own AutomationBench row. After November 30 it costs the same as Claude Opus 5.5, which beats it on several coding tests, so test both.

Gemini 4 Argon: the in-between

Argon handles long-horizon tasks with fewer tokens than Carbon. With Hydrogen now matching its index score, Argon’s niche has narrowed: pick it when you need long-context strength (it led GraphWalks 256K–1M at launch) and image generation, but not Carbon’s depth. Our Argon launch analysis has the full benchmark set.

flowchart TD classDef q fill:#f1f5f9,stroke:#64748b,color:#0f172a classDef g fill:#dbeafe,stroke:#2563eb,color:#1e3a8a S["New request"]:::q --> H["Try Gemini 4 Hydrogen\n(high reasoning)"]:::g H --> Q1{"Good enough?"}:::q Q1 -->|yes| D["Done"]:::q Q1 -->|no| X["Hydrogen at xhigh"]:::g X --> Q2{"Good enough?"}:::q Q2 -->|yes| D Q2 -->|"no: very long input\nor needs images"| A["Gemini 4 Argon"]:::g Q2 -->|"no: hard reasoning\nor security"| C["Gemini 4 Carbon"]:::g

Escalate only when the cheaper setting fails

The December 1 price change

Carbon and Argon launched at introductory prices. From December 1, 2026 both double:

ModelUntil Nov 30From Dec 1Cost of 10M in + 2M out per day, 30 days
Gemini 4 Carbon$2 / $10$4 / $20$1,200 › $2,400
Gemini 4 Argon$2 / $10$4 / $20$1,200 › $2,400
Gemini 4 Hydrogen$1.50 / $7.50$1.50 / $7.50$900
Gemini 3.8 Flash$0.75 / $3.75$0.75 / $3.75 (rises to $1.50 / $7.50 on Jan 1, 2027)$450 › $900

If you built on Carbon or Argon during the intro period, re-run your evals on Hydrogen at xhigh before December. Many workloads will pass at roughly a third of Carbon’s new price. Flash users face the same decision a month later, when 3.8 Flash rises to Hydrogen’s price.

Combining Carbon and Hydrogen

Google’s Antigravity Orchestrator switches between Carbon and Hydrogen in one session: Carbon plans, Hydrogen executes. Google reports this setup came close to Carbon’s Terminal-Bench 4.0 success rate while finishing 19% faster with 24% less quota. You can build the same pattern on the API: a planner call to Carbon, then many Hydrogen calls for the steps. Our routing guide shows how.

Gemini 4 vs Claude and GPT-6.2

TierGoogleAnthropicOpenAI
TopCarbon: AA 60, $4/$20 from Dec 1Fable 5.2: AA #1 (65), $8/$40GPT-6.2 Astra: AA #2, $8/$40
Strong valueArgon: AA 50, $4/$20 from Dec 1Opus 5.5: AA 59, $4/$20—
EverydayHydrogen: AA 51–54, $1.50/$7.50Sonnet 5.5: AA 55, $2/$10GPT-6.2 Sol: AA 57, $2/$10
Cheapest3.8 Flash: $0.75/$3.75Haiku 5.5: $0.10/$0.50GPT-6 Luna: $0.10/$0.50

Google’s strengths are price at the everyday tier, the free app, blind-vote preference and integration with Workspace and Search. At the very top, Claude Fable 5.2 and GPT-6.2 Astra still score higher on independent tests.

FAQ

What is the difference between Gemini 4 Carbon, Argon and Hydrogen?

They share one architecture and a 1,048,576-token context window but spend different amounts of thinking per turn. Carbon is the deepest reasoner (AA Index 60), Argon handles long-horizon tasks with fewer tokens, and Hydrogen is the fast, cheap successor to Gemini 3.8 Flash.

How much does Gemini 4 Hydrogen cost?

$1.50 per million input tokens and $7.50 per million output tokens, with cached input at $0.075. Batch and Flex are half price, and there is a free API tier. Free Gemini app users also get Hydrogen through automatic model selection.

Is Gemini 4 Carbon getting more expensive?

Yes. Carbon and Argon cost $2/$10 per million tokens through November 30, 2026, then $4/$20 from December 1.

Is Gemini 4 Hydrogen better than Argon?

On the Artificial Analysis Intelligence Index, slightly: Hydrogen scores 51 at high reasoning and 54 at xhigh, against Argon’s 50. Argon still has image generation and Live API support, which Hydrogen lacks, and ranks higher in blind user votes.

Is Gemini 4 better than Claude?

At the top, no: Claude Fable 5.2 leads the AA Index and most coding tests. Gemini 4 Carbon wins blind-vote preference on LMArena and AutomationBench, and Gemini 4 Hydrogen is cheaper than Claude Sonnet 5.5 for everyday work.

Sources

Prices and scores as published October 28 – November 18, 2026. Check Google’s Gemini API pricing page before budgeting.

Building on Gemini?

stackcone builds agents on Gemini, Claude and GPT models with routing between them, so you pay Carbon prices only for Carbon-sized problems.