Which Gemini 4 Model? Argon, Carbon, Hydrogen, Neon vs Claude

Google says to default to Gemini 4 Hydrogen ($1.50/$7.50 per million tokens, the only one with a free tier), use Neon ($3/$15) for complex tasks and Carbon for the hardest problems. Argon and Carbon rise from $2/$10 to $4/$20 on 1 December 2026. Three Claude models still lead Arena WebDev. Checked 20 November 2026.

Viết với sự hỗ trợ của AI: bản nháp hướng dẫn này được soạn cùng Claude Sonnet 5.5, và mỗi dữ kiện đều dẫn link đến nguồn của nó.

Google's advice is to default to Gemini 4 Hydrogen, use Neon for complex, high-frequency tasks and keep Carbon for the hardest problems (choosing a model, checked 20 November 2026). For animation and web pages, Claude Fable 5.2, Opus 5.5 and Sonnet 5.5 rank first to third on Arena WebDev; Gemini 4 costs less per run.

What are Gemini 4 Argon, Carbon, Hydrogen and Neon?

They are Google's four Gemini 4 models, made generally available between 13 October and 18 November 2026. All four read text, images, audio, video and PDFs, write only text, and share one 1,048,576-token window for input and output, with 65,536 output tokens per normal request (Carbon model page). Their knowledge cutoff is March 2026 (Argon model card).

The Gemini 4 lineup from Google's API docs, checked 2026-11-20
Model and API IDGA datePrice per 1M tokens now (in / out)From 2026-12-01Default thinking (levels)Context / max outputWhere it runs
Hydrogen, gemini-4-hydrogen12 Nov 2026$1.50 / $7.50, plus a free tierNo changehigh (low, medium, high, xhigh)1M shared / 65,536Free and paid API, AI Studio, Agent Platform, Gemini app (all plans), Antigravity (paid Pro, Ultra)
Neon, gemini-4-neon18 Nov 2026$3 / $15No changehigh (low, medium, high, xhigh)1M shared / 65,536Paid API, AI Studio, Agent Platform (verified enterprise), Gemini app and Antigravity (Pro, Ultra)
Carbon, gemini-4-carbon10 Nov 2026$2 / $10$4 / $20medium (low, medium, high)1M shared / 65,536Paid API, AI Studio, Agent Platform (verified enterprise), Gemini app and Antigravity (Ultra only)
Argon, gemini-4-argon13 Oct 2026$2 / $10$4 / $20medium (low, medium, high)1M shared / 65,536Paid API, AI Studio, Agent Platform, Antigravity (Pro, Ultra); left the Gemini app 18 Nov
SourceDeprecations tablePricingPricingThinking docsCarbon model pageGemini app help, Antigravity models

Which Gemini 4 model should you use for which job?

Google's choosing-a-model page sets the order as of 20 November 2026: default to Hydrogen, use Neon for complex, high-frequency tasks and Carbon for your hardest problems. Argon is no longer on that page. The Gemini API has no automatic router, so you name the model on each call; Antigravity's Orchestrator plans with Neon, Carbon or Argon and implements with Hydrogen.

motif's reading for animation and web work, from the boards below:

Which Gemini app plan includes which Gemini 4 model?

As of 20 November 2026 the Gemini app gives Hydrogen to every plan, Neon to AI Pro and Ultra, and Carbon to AI Ultra only, according to Google's Gemini models help page. No plan lists Argon. US prices are $4.99, $19.99 and $99.99 or $199.99 a month (Google AI plans).

Gemini 4 models by Google AI plan, personal accounts, checked 2026-11-20
Plan (US price a month)Gemini app chatsGemini app TasksAntigravityAI Studio
Free ($0)Auto only (Hydrogen or a lightweight model)HydrogenNo Gemini 4 modelHydrogen
AI Plus ($4.99)Auto, HydrogenHydrogenNo Gemini 4 modelHydrogen
AI Pro ($19.99)Auto, Hydrogen, NeonHydrogen, NeonHydrogen, Neon, Argon (paid Pro)Hydrogen, Neon, Argon
AI Ultra ($99.99 or $199.99)Auto, Hydrogen, Neon, CarbonHydrogen, Neon, CarbonAll fourAll four
SourceGemini app helpGemini app helpAntigravity modelsAI Studio plans

Usage limits are compute-based, not per model, and refresh every 5 hours up to a weekly cap: Plus gets 2x the standard limit, Pro 4x and Ultra 5x or 20x Pro (Gemini limits).

On Claude's side, Free includes Sonnet but not Opus or Fable, and Pro at $20 a month adds Opus (Claude pricing).

How much does one animation or web page cost on each model?

For a sample job of 3,000 input and 12,000 output tokens, roughly one single-file animation or landing page, Gemini 4 Hydrogen costs about $0.09, Neon $0.19, and Carbon or Argon $0.13 until 30 November 2026 and $0.25 from 1 December. Claude Sonnet 5.5 costs $0.13, Opus 5.5 $0.25 and Fable 5.2 $0.50 (Gemini pricing, Claude pricing).

Cost of one 3k-input, 12k-output generation at standard API prices, checked 2026-11-20
ModelPrice per 1M tokens (in / out)ArithmeticNowFrom 2026-12-01Source
Gemini 4 Hydrogen$1.50 / $7.500.003 x $1.50 + 0.012 x $7.50$0.0945$0.0945Gemini API pricing
Gemini 4 Neon$3 / $150.003 x $3 + 0.012 x $15$0.189$0.189Gemini API pricing
Gemini 4 Carbon or Argon$2 / $10, then $4 / $200.003 x $2 + 0.012 x $10, then 0.003 x $4 + 0.012 x $20$0.126$0.252Gemini API pricing
Claude Sonnet 5.5$2 / $100.003 x $2 + 0.012 x $10$0.126$0.126Claude pricing
Claude Opus 5.5$4 / $200.003 x $4 + 0.012 x $20$0.252$0.252Claude pricing
Claude Fable 5.2$8 / $400.003 x $8 + 0.012 x $40$0.504$0.504Claude pricing

Gemini 4 output prices include thinking tokens, and Batch or Flex halves them. Real runs vary: Simon Willison's same SVG prompt at high thinking used 20,305 output tokens (20.31 cents) on Carbon and 5,311 (3.99 cents) on Hydrogen (his notes).

Is Gemini 4 Argon being replaced?

Partly. Argon left the Gemini app on 18 November 2026, when Google's plan page swapped it for Neon, and AI Mode dropped it when Carbon launched. In the Gemini API, gemini-4-argon has no shutdown date (deprecations), and its price doubles to $4/$20 on 1 December 2026.

The Agent Platform retirement date, GitHub Copilot's 14 December cut-off and every place Argon can still be picked are in the Argon status guide.

How do the Gemini 4 models score on independent benchmarks?

At high thinking, Carbon and Neon lead the family. Carbon scores 60.03 and Neon 59.62 on the Artificial Analysis index, and Gemini 4 holds places 1 to 3 on Arena Text but peaks at 4th on WebDev. Claude Fable 5.2 at max leads the whole AA index at 65 (AA leaderboard), checked 20 November 2026.

Gemini 4 at high thinking on independent boards, checked 2026-11-20
Model (High)AA Index v5.0.1AA Terminal-Bench 4.0Arena Text (of 426)Arena WebDev (of 162)Arena Agent (of 68)Vals Index v2.2
Carbon60.0365.15%1,527 (#1)1,769 (#4)1,303 (#4)70.45% (#4)
Neon59.6261.62%1,513 (#3)1,763 (#5)1,333 (#3, preliminary)70.87% (#2, run at xhigh)
Hydrogen50.7954.04%1,490 (#18)1,701 (#14)1,271 (#10, preliminary)68.49% (#7, run at xhigh)
Argon50.4157.07%1,520 (#2)1,681 (#18)1,229 (#14)69.39% (#6)
SourceAA model pages: Carbon, Neon, Hydrogen, ArgonSame AA pages (per-level data embedded in each page)Arena TextArena WebDevAgent ArenaVals Index

Claude leads Agent Arena, with Fable 5.2 at 1,448 and Opus 5.5 at 1,336, and the Vals Index, with Fable 5.2 at 71.31% (Vals Index); Claude entries run at max or xhigh. Google's own model cards are not comparable with these boards: Gemini runs at its highest thinking setting and rival figures are mostly the providers' own (Neon evaluation notes).

Is Gemini 4 or Claude better for animation and web design?

Claude, on most human-vote boards checked 20 November 2026. Fable 5.2 (1,823), Opus 5.5 (1,818) and Sonnet 5.5 (1,771) rank first to third on Arena WebDev, and Opus 5.5 leads Design Arena's agentic board. Gemini 4 beats Claude on Design Arena's SVG board, Arena's HTML category and cost, and Neon nearly ties Fable 5.2 on Vibe Code Bench.

Gemini 4 against Claude on web, design and coding-agent boards, checked 2026-11-20
BoardBest ClaudeBest Gemini 4Leader
Arena WebDev, overallFable 5.2 xHigh, 1,823 (#1)Carbon High, 1,769 (#4)Claude (Carbon and Neon overlap Sonnet 5.5)
Arena WebDev HTMLFable 5.2 xHigh, 1,778 (#6)Neon High, 1,788 (#5)Neon, ahead of all three Claude models
Vals Vibe Code Bench v1.1Fable 5.2, 93.77% (#1, $22.13 per test)Neon, 93.70% (#2, $6.76 per test)Near tie; Neon costs about a third
Design Arena, overall agenticOpus 5.5, 1,442 (#1)Carbon, 1,304 (#9)Claude
Design Arena, combinedOpus 5.5, 1,378 (#5)Neon, 1,394 (#1, on 1,239 battles)Neon, on a small sample
Design Arena SVGOpus 5.5, 1,316 (#8)Carbon, 1,394 (#3)Gemini 4
AA Coding Agent Index v2.1Fable 5.2 xhigh (Claude Code) and Opus 5.5 max, 58Neon xhigh, 59 (Antigravity CLI)Neon by one point
AA-3D, Three.js scenesFable 5.2, 1,271Argon High, 1,147 (only Gemini 4 model listed)Claude

Antigravity offers all four Gemini 4 models on Ultra, plus Claude Opus 5.5 and Sonnet 5.5 from paid AI Pro up, and Google's Carbon post claims "improved front end taste and motion design skills" there. Claude Code defaults to Opus 5.5, and Claude Motion, a Team and Enterprise beta, renders code animations to MP4 (explained).

Every model here outputs text only, so an animation is HTML, CSS, SVG or Three.js code that you render to video afterwards (HTML to MP4). To choose between Claude models, see Opus 5.5 vs Sonnet 5.5 for animation.

What have creators found when testing Gemini 4 against Claude?

No board compares animation output yet: Design Arena's Code to Video board was still "coming soon" on 20 November 2026 (Design Arena). The creator tests published so far mostly favour Claude on output quality and Gemini 4 on cost. Each is one tester's result, not a controlled benchmark.

motif's archive of 28 Gemini 4 posts, checked 20 November 2026, has none on Carbon, Hydrogen or Neon yet; its 22 same-prompt Argon runs against Claude and the creators' verdicts are in the Argon guide.

How do you use Gemini 4 in Antigravity, Gemini CLI, Copilot, Cursor and OpenRouter?

Antigravity is Google's route to the full lineup: on 20 November 2026 it offered Hydrogen, Neon and Argon on paid AI Pro, all four on Ultra and none on Free or Plus (Antigravity models). Neon has been its default since 18 November (changelog). GitHub Copilot, Cursor and OpenRouter list all four.

  • Gemini CLI: not an official route; its README says "Not tuned for the Gemini 4 series", and Google moved AI Pro and Ultra users to Antigravity CLI on 18 June 2026.
  • GitHub Copilot: Hydrogen from Copilot Pro up; Argon, Carbon and Neon need Pro+, Max, Business or Enterprise (GitHub docs).
  • Cursor: all four at Google's prices on Premium, Super or Ultra, not Free or Plus (Cursor docs).
  • OpenRouter: google/gemini-4-argon, -carbon, -hydrogen and -neon at Google's prices (Neon on OpenRouter).

Why are the Gemini 4 models named after elements?

Google has not said, as of 20 November 2026. Its Neon post pairs Bunsen flames with Carbon and Hydrogen and gas discharge tubes with Argon and Neon, but release order follows neither atomic number nor capability. The family has no Pro model and no link to NASA's 1965 Gemini 4 flight. Argon's launch post called it "just the beginning".

Questions

Which Gemini 4 model should I use?

Google's default is Hydrogen ($1.50/$7.50 per million tokens), with Neon ($3/$15) for complex, high-frequency tasks and Carbon for the hardest problems, checked 20 November 2026. Carbon scores highest of the four on Artificial Analysis at high thinking (60.03), with Neon close at 59.62.

Is Gemini 4 better than Claude for animation and web design?

Not on most boards checked 20 November 2026. Claude Fable 5.2, Opus 5.5 and Sonnet 5.5 rank first to third on Arena WebDev, ahead of Carbon (4th) and Neon (5th). Gemini 4 leads Design Arena's SVG board and costs less per run.

How much does Gemini 4 cost per million tokens?

Hydrogen costs $1.50 input and $7.50 output and Neon $3/$15, with no change on 1 December 2026. Argon and Carbon cost $2/$10 through 30 November 2026 and $4/$20 from 1 December. Batch and Flex are half price, checked 20 November 2026.

Is Gemini 4 free?

Only Hydrogen. It is the one Gemini 4 model on the free Gemini API tier, and free Gemini app users get it through Auto. Argon, Carbon and Neon are paid only, from $2/$10 per million tokens, checked 20 November 2026.

What are the Gemini 4 model IDs?

gemini-4-argon, gemini-4-carbon, gemini-4-hydrogen and gemini-4-neon. All four are stable IDs with no preview, dated or -latest alias, per Google's API docs checked 20 November 2026. OpenRouter uses google/gemini-4-neon and the matching slugs.

What are the Gemini 4 context window and knowledge cutoff?

All four share one 1,048,576-token window for input and output, with up to 65,536 output tokens per normal request, and a March 2026 knowledge cutoff. Claude Fable 5.2, Opus 5.5 and Sonnet 5.5 have 1M context and 128K output per request, checked 20 November 2026.

Is Gemini 4 Argon being replaced?

Partly. Argon left the Gemini app on 18 November 2026 and AI Mode when Carbon launched, but the Gemini API lists no shutdown date. Its price doubles to $4/$20 on 1 December 2026, and GitHub Copilot drops it on 14 December 2026.

Which Google AI plan includes which Gemini 4 model?

As of 20 November 2026 the Gemini app gives Hydrogen to every plan, Neon to AI Pro ($19.99 a month in the US) and Ultra, and Carbon to AI Ultra only ($99.99 or $199.99). No app plan lists Argon. Claude's Free plan includes Sonnet; Pro at $20 adds Opus.

Can Gemini 4 make animations or videos?

It writes the code for them. All four Gemini 4 models output only text, so an animation is HTML, CSS, SVG, GSAP, Three.js or WebGL2 code that runs in a browser and is rendered to MP4 afterwards, the same as with Claude (checked 20 November 2026).

Why is there no Gemini 4 Pro?

Google has released no model called Gemini 4 Pro. As of 20 November 2026 the family is Argon, Carbon, Hydrogen and Neon, and Google's recommended upgrade from Gemini 3.1 Pro Preview is Neon, or Carbon for the hardest problems.

Nguồn

  1. Google AI Studio docs: Choosing a model
  2. Google AI Studio docs: Gemini API pricing
  3. Google AI Studio docs: Gemini 4 Carbon model page
  4. Google AI Studio docs: Thinking
  5. Google AI Studio docs: Deprecations
  6. Google AI Studio docs: Google AI plans
  7. Google: Introducing Gemini 4 Argon
  8. Google: Gemini 4 Carbon is now available
  9. Google: Introducing Gemini 4 Neon
  10. Google DeepMind: Gemini 4 Argon model card
  11. Google DeepMind: Gemini 4 Neon evaluation methodology (PDF)
  12. Gemini Apps Help: About Gemini models
  13. Gemini Apps Help: Usage limits
  14. Google AI plans (US)
  15. Wayback Machine: Google AI plans page, 18 November 2026
  16. Google Search Help: AI Mode models
  17. Google Antigravity docs: Models
  18. Google Antigravity docs: Changelog
  19. google-gemini/gemini-cli README
  20. Claude Platform Docs: Pricing
  21. Claude pricing and plans
  22. Claude Code Docs: Model configuration
  23. Claude Help Center: Get started with Claude Motion
  24. GitHub Docs: Supported AI models in Copilot
  25. GitHub Changelog: Upcoming deprecation of selected Copilot models in December
  26. Cursor Docs: Models and pricing
  27. OpenRouter: Gemini 4 Neon
  28. Artificial Analysis: Models leaderboard
  29. Artificial Analysis: Gemini 4 Carbon
  30. Artificial Analysis: Gemini 4 Neon
  31. Artificial Analysis: Gemini 4 Hydrogen
  32. Artificial Analysis: Gemini 4 Argon
  33. Artificial Analysis: Gemini 4 Neon intelligence gain (Coding Agent Index)
  34. Artificial Analysis: AA-3D
  35. Arena: Text leaderboard
  36. Arena: WebDev leaderboard
  37. Arena: WebDev HTML category
  38. Arena: Agent Arena leaderboard
  39. Vals AI: Vals Index
  40. Vals AI: Vibe Code Bench
  41. Design Arena: Leaderboard
  42. Design Arena: Combined overall board
  43. Design Arena: SVG board
  44. Simon Willison: Gemini 4 models in LLM
  45. Simon Willison: yonkers-rising posts (Three.js mech prompt)
  46. AI for Mortals (Pat Simmons): Gemini 4 Carbon review
  47. Peter Yang: Gemini 4 Neon vs Claude Opus 5.5
  48. AICodeKing: Gemini 4 Carbon vs Claude Fable 5.2
Tất cả hướng dẫn