Blog · AI models

Gemini 4 Models Explained: Carbon, Argon, Neon and Hydrogen Compared

Google has released four Gemini 4 models: Carbon, its strongest model, for the hardest long-running work; Argon, the first Gemini 4 model, an efficient frontier reasoning model; Neon, the newest, which "balances depth and efficiency"; and Hydrogen, the fast and cheapest one, and the only one with a free API tier. All four have a 1M-token context and are named after chemical elements. This guide compares their release dates, prices, availability and benchmarks, gives Google's own advice on which to pick, and sorts out the names that aren't real models. Updated November 21, 2026.

Gemini 4 release dates

Google announced the two larger models to cyber defenders first and opened them to developers about two weeks later. Hydrogen and Neon were public on the day they were announced. Gemini 4 came out in this order:

  • Argon: announced September 30, 2026 for "a set of trusted cyber defenders through our Fairwind Program" (Google), and broadly available on October 13, 2026 in the Gemini API, AI Studio, Antigravity and the Gemini app (Google). More in our Gemini 4 Argon guide.
  • Carbon: announced October 28, 2026 for "a subset of cyber defenders" (Google), and broadly available on November 10, 2026 (Google). More in our Gemini 4 Carbon guide.
  • Hydrogen: released November 12, 2026, for everyone in the Gemini app and on every AI Studio tier (Google). More in our Gemini 4 Hydrogen guide.
  • Neon: released November 18, 2026 (Google). More in our Gemini 4 Neon guide.
  • Some sites date Argon September 30 and Carbon October 28. Those are the limited announcements; public API access started October 13 and November 10.

Gemini 4 Carbon vs Argon vs Neon vs Hydrogen

Prices are Gemini API Standard rates per million tokens from Google's pricing page on November 21, 2026. Output prices include thinking tokens. Carbon and Argon have a launch price through November 30, 2026 and a higher price from December 1. Neon and Hydrogen have no dated price change. All four models have a 1,048,576-token context window shared by input and output, and return up to 65,536 tokens per streamed segment.

CarbonArgonNeonHydrogen
Public releaseNov 10, 2026 (announced Oct 28)Oct 13, 2026 (announced Sep 30)Nov 18, 2026Nov 12, 2026
What Google says it's for"our strongest model""Built to solve your hardest problems"; "efficient frontier reasoning model""balances depth and efficiency""our daily driver for agentic workflows"
Model IDgemini-4-carbongemini-4-argongemini-4-neongemini-4-hydrogen
Price in / out (through Nov 30)$2 / $10$2 / $10$3 / $15$1.50 / $7.50
Price in / out (from Dec 1)$4 / $20$4 / $20$3 / $15$1.50 / $7.50
Cached input$0.10, then $0.20$0.10, then $0.20$0.15$0.075
Free API tierNoNoNoYes
Thinking levelslow, medium (default), highlow, medium (default), highlow, medium, high (default), xhighlow, medium, high (default), xhigh
Gemini appAI Ultra (rolling out)AI Pro and Ultra (from Oct 13)AI Pro and UltraEveryone (free users through auto model selection)
Google AntigravityAI UltraAI UltraPaid AI Pro and UltraPaid AI Pro and Ultra
AI Studio and Gemini APIPaid tierPaid tierPaid tierAll tiers, including free

Which Gemini 4 model to pick: Google's guide

Google's AI Studio page Choosing a model ranks three of the four by speed and says what each is for. Argon isn't on that page.

  • Hydrogen (fast): "interactive web apps, routine code edits, multimodal document pipelines, and parallel worker subagents".
  • Neon (moderate): "multi-file refactoring, repository-wide debugging, and analytical tasks", when someone is waiting on the answer.
  • Carbon (deliberate): "multi-hour background runs, deep research, formal proofs, and security audits". Google says to move up to Carbon when Neon at high or xhigh gets stuck.
  • Before switching models, Google suggests tuning thinking_level first. You can also plan with Carbon or Neon and implement with Hydrogen, passing context along with previous_interaction_id.
  • Moving from older models: Google points gemini-2.5-pro and gemini-3.1-pro-preview users to Neon or Carbon, and most Flash-tier users to Hydrogen.
  • Argon: Google's models page calls it an "Efficient frontier reasoning model with faster responses and lower token spend per turn than Carbon". Google Cloud's Neon developer guide calls Neon "the target upgrade" from Argon.

Gemini 4 Carbon vs Neon vs Argon

These three are the larger Gemini 4 models. Carbon is the strongest and slowest, Neon is the newest and sits in the middle, and Argon came first.

  • Carbon vs Neon: Google keeps Carbon for multi-hour background runs and Neon for complex work where someone is waiting. On Artificial Analysis, Neon scores 61 (xhigh) on the Intelligence Index and Carbon 60 (high); on AA's Coding Agent Index, Neon 59 and Carbon 55. Neon costs more per token than Carbon through November 30 and 25% less from December 1 ($3 / $15 against $4 / $20).
  • Argon vs Carbon: both cost the same and both default to medium thinking. Google says Carbon uses more tokens than Argon to explore and verify. In its Carbon launch post on November 10, before Neon existed, Google said Argon "remains the ideal choice for the majority of software engineering and knowledge work" and Carbon is for harder, open-ended tasks. AA scores Carbon 60 and Argon 50.
  • Argon vs Neon: Google Cloud names Neon the upgrade from Argon. Neon defaults to high thinking and adds xhigh, where Argon defaults to medium and tops out at high. Neon also takes Office and OpenDocument files on Google Cloud, while Argon's europe-west2 and asia-southeast1 Provisioned Throughput locations aren't supported for Neon. AA scores Neon 61 and Argon 50. On Arena's text board (November 20), Argon ranks one place above Neon.
  • Argon is still listed and priced in the Gemini API, and we found no deprecation notice.

Gemini 4 benchmarks Google published

These are Google's own numbers, from each model's launch post, on benchmarks where Google gave a score for more than one Gemini 4 model. The posts came out on different dates, so treat any "#1" as of that date. A dash means Google didn't publish a number for that model.

Benchmark (Google's numbers)CarbonArgonNeonHydrogen
Terminal-Bench 4.1--65.2%57%
HLE-Diamond--71.3%58.8%
Agents' Last Exam v1.1--75.8%71.6%
SEC-bench Pro88.2%-90.5%-
Zapier AutomationBench#1 (no score given)51.3%, #1 on Sep 3056.3%, #1-
Gray Swan prompt-injection attack success (lower is better)0.0%0.7%--

Other headline numbers, per Google

  • Carbon (Oct 28, Nov 10): CWE-bench 71.0%, ExploitGym 54.8%, WebDev Arena #4 at 1768 Elo, agentic honesty 92% (Argon 70%). Google also claims #1 on CursorBench 4.0 and Perplexity WANDR without giving scores.
  • Argon (Sep 30, Oct 13): DeepSWE v1.1 77.9%, LVBench 91.7%, CWE-bench v1 68% (tied #1), ERQA 87%, and #1 on the Vals Index (no score given).
  • Neon (Nov 18): Vals Index 70.9% (#2), FrontierCode 2.0 46.2%, Agent Arena #2 at 1336. See Gemini 4 Neon for the full list.
  • Hydrogen (Nov 12): Vals Tax Agent Bench 52.3% (#1), Vals Legal Research Bench 55.8% (tied #1), MKA 53.4%, about 2.5x faster than Gemini 3.8 Flash. See Gemini 4 Hydrogen.

Gemini 4 on Artificial Analysis and Arena (third party)

Artificial Analysis (AA) runs its own tests on each model. Arena (formerly LMArena) ranks models from people's votes. Ranks change quickly: Google quoted Carbon as #1 on AA with 59 on October 28; on November 21 AA shows it at 60 and #8. Each figure below carries the date we read it.

CarbonArgonNeonHydrogen
AA Intelligence Index60 (high)50 (high)61 (xhigh)54 (xhigh), 51 (high)
AA rank#8 of 95 (Nov 21)#33 of 95 (Nov 21)#6 of 96 (Nov 19)#26 of 95 (Nov 21)
AA Coding Agent Index v2.155 (high)-59 (xhigh)-
AA cost per Intelligence Index task$7.76$4.97$6.44$3.16
AA output speed55 tokens/s69 tokens/sabout 110 tokens/sabout 161-166 tokens/s
Arena Text rank (Nov 20)#1, 1527#2, 1520#3, 1513Not ranked

Where the third-party numbers come from

  • AA model pages: Carbon, Argon, Hydrogen, and AA's Neon article for Neon and the Coding Agent Index.
  • AA's cost per task for Carbon uses its post-promo $4 / $20 price.
  • Arena's text leaderboard, overall view, dated November 20, 2026. Gemini 4 models hold the top three places. Hydrogen appears in Arena's chart but has no rank on the board.

Is Gemini 4 Oxygen, Krypton, Sodium or Pro real?

No, not as of November 21, 2026. Google's models and pricing pages list exactly four Gemini 4 models: Argon, Carbon, Hydrogen and Neon.

  • Gemini 4 Pro: not an official name. Google's "Pro" models are older ones such as Gemini 3.1 Pro and 2.5 Pro. "Gemini 4 Pro" shows up in speculation from before the launch and on third-party pages. If you used 2.5 Pro or 3.1 Pro, Google points you to Neon or Carbon.
  • Gemini 4 Krypton: a rumor. It started with an X post by @lyraxana on November 8, 2026 claiming Google had an internal checkpoint called Krypton that beat Carbon, followed by comparison posts on November 9. OfficeChai summarized them and notes that Google hasn't confirmed it. Krypton isn't on any Google page, Artificial Analysis or Arena.
  • Gemini 4 Oxygen: a rumor. The name appeared on X on November 15, 2026, a day after an unlabeled model showed up on Arena, and testers describe it as a fast, Flash-class model. Kie.ai collected the posts and notes that Google and DeepMind haven't mentioned it. Oxygen isn't on any Google page.
  • Gemini 4 Sodium: we found no evidence of it. Searches return only NASA's 1965 Gemini 4 mission, and it isn't on Google's model or pricing pages.
  • Gemini 4 Flash: also unconfirmed. Google's guide sends Flash-tier users to Hydrogen.

How to try the Gemini 4 models

  1. Start free with Hydrogen

    Hydrogen is on every AI Studio tier, including free, and in the Gemini app for everyone. Carbon, Argon and Neon need a paid API key.

  2. Call a model from the Gemini API

    Swap in the model ID you want. Tune thinking_level before switching to a bigger model.

    # Gemini API model IDs
    # gemini-4-hydrogen  (free tier available)
    # gemini-4-neon      (paid tier)
    # gemini-4-argon     (paid tier)
    # gemini-4-carbon    (paid tier)
    curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-4-hydrogen:generateContent" \
      -H "x-goog-api-key: $GEMINI_API_KEY" \
      -H "Content-Type: application/json" \
      -d '{"contents": [{"parts": [{"text": "Find and fix the failing test in this diff"}]}]}'
  3. Use them in the Gemini app

    Hydrogen is there for everyone. AI Pro and Ultra subscribers can pick Neon in the model dropdown, and Carbon is rolling out to AI Ultra.

  4. Code with them in Google Antigravity

    Hydrogen and Neon are on paid AI Pro and Ultra plans; Carbon and Argon need AI Ultra.

Source: Google AI Studio, Choosing a model

Games made with Gemini 4

The first Neon game you can play is NEON SCRAMBLE, which Google posted on November 19: a 3D runner where you guide Google Maps' Pegman from Shibuya Scramble Crossing to the Google building in Tokyo. Play it or see it on vibegamer.io.

Games built with Gemini models are collected on vibegamer.io: made with Gemini, Gemini 4 Neon games and Gemini 4 Hydrogen games.

With Crewmark

A new model is half of it. The other half is a job.

Models keep getting better and cheaper, and the agents around them still wait for the next prompt. Crewmark turns coding agent sessions into AI employees with a role, a goal in numbers, working hours and a heartbeat, so they take the next issue from a board on their own. Claude Code and Codex employees run natively; other agents, such as OpenCode, DeepSeek Harness and Gemini, connect through ACP (acpx). An employee uses whatever model its agent runs.

  1. Pick a role

    From the AI employee roles, such as an AI software engineer.

  2. Put the work on a board

    Assign issues, make the regular ones repeat, and let the heartbeat wake the employee to take the next one.

  3. See what it can do

    Jobs with setup and rules for each agent are in Claude Code use cases and Codex use cases.

FAQ

What are the Gemini 4 models?

Google has released four: Carbon, its strongest model; Argon, an efficient frontier reasoning model and the first Gemini 4 release; Neon, which balances depth and efficiency; and Hydrogen, the fast, cheapest model with a free API tier. All four have a 1M-token context.

When was Gemini 4 released?

Argon was announced September 30, 2026 and opened broadly on October 13. Carbon was announced October 28 and opened broadly on November 10. Hydrogen came out November 12 and Neon November 18, 2026.

Gemini 4 Carbon vs Neon: which should I use?

Google's guide says Neon for complex coding and analysis when someone is waiting, and Carbon for multi-hour background runs, deep research, proofs and security audits, or when Neon gets stuck. Artificial Analysis scores them close (Neon 61, Carbon 60). From December 1, Neon is 25% cheaper per token.

Gemini 4 Argon vs Neon: what's the difference?

Google Cloud calls Neon the target upgrade from Argon. Neon defaults to high thinking and adds xhigh; Argon defaults to medium. Neon costs $3 / $15 per million tokens; Argon is $2 / $10 through November 30 and $4 / $20 from December 1.

Gemini 4 Argon vs Carbon: which is better?

Carbon is Google's strongest model and uses more tokens to explore and verify. Both cost the same. Artificial Analysis scores Carbon 60 and Argon 50; on Arena's text board (November 20) Carbon is #1 and Argon #2.

Is there a Gemini 4 Pro?

No. Gemini 4 Pro isn't an official model name; Google's Gemini 4 models are Argon, Carbon, Hydrogen and Neon. Google points Gemini 2.5 Pro and 3.1 Pro users to Neon or Carbon.

Is Gemini 4 Krypton or Sodium real?

Neither is confirmed. Krypton is a rumor that started with an X post on November 8, 2026; Google hasn't confirmed it. We found no evidence of a Gemini 4 Sodium at all.

Which Gemini 4 model is free?

Hydrogen. It has a free tier in the Gemini API and AI Studio and is in the Gemini app for everyone. Carbon, Argon and Neon need a paid API tier or a Google AI Pro or Ultra plan.

Checked against Google's launch posts for each model, the Gemini API pricing, models and choosing-a-model pages, Google Cloud's model docs, Artificial Analysis and Arena on November 21, 2026. Some Neon details are from our November 19 check.

Put the newest models on a standing job.

Crewmark turns coding agent sessions into AI employees with roles, goals, working hours and a heartbeat.

Crewmark is open source: star us on GitHub