AI

Gemini 4 Carbon vs. Hydrogen: Deep Reasoning or Fast Execution?

Google's two newest models share a lineage but are built for different jobs, and their price gap is set to widen on December 1.

11 min read
Long, dim server room with rows of black server racks, small blue indicator lights, and ceiling lights receding into the distance.
Image by Schäferle from Pixabay

Carbon is Google's strongest Gemini 4 model, aimed at hard, long-horizon reasoning. Hydrogen, released November 12, 2026, is a faster, cheaper model adapted from Carbon for agent workflows. Most performance evidence is vendor-reported or comes from commercial benchmark services, not peer review.

Google's two newest Gemini 4 models are built for different kinds of work. Gemini 4 Carbon, announced October 28 and made generally available November 10, 2026, is the company's strongest model and targets hard, long-horizon reasoning. Gemini 4 Hydrogen, released November 12, 2026, is a faster and cheaper model adapted from Carbon, designed for rapid execution in agentic workflows, meaning software that plans and carries out multistep tasks with little supervision.

The choice matters for cost and access as much as for quality. Carbon's introductory API price matches that of Google's Argon model, but it is reported to double on December 1, while Hydrogen's price is flat. Plan tiers also decide who can use which model. The performance evidence is young: it is vendor-reported or comes from commercial benchmark services, and none of it has been peer reviewed.

How the two models are related

Release dates

Hydrogen's release date is November 12, 2026, not November 13. Google Cloud's model documentation lists general availability and a November 12 release date, and the independent benchmarker Artificial Analysis gives the same date. A launch tracker that cites Google's announcement, titled "Introducing Gemini 4 Hydrogen," timestamps it at 18:00 UTC on November 12, which is still that day in every US time zone.

Where November 13 appears, it marks when an outlet published or checked its coverage. MarkTechPost's article, for example, carries a 2026/11/13 date in its address, and one tracker lists November 13 as its own verification date while giving November 12 as the launch.

Carbon's timeline has two dates. Google announced it on October 28, 2026, and said it would reach developers, businesses and consumers "soon." Google's availability post followed on November 10. Independent outlets agree on both dates, and one notes the 13-day gap matches the one between Argon's announcement and its rollout. Argon, the third Gemini 4 model, reached general availability on October 13, according to the brief's sources.

Lineage and positioning

Google's own documentation states the relationship plainly. The Hydrogen model card says "Gemini 4 Hydrogen is adapted from Gemini 4 Carbon." Secondary sources say Carbon in turn builds on Argon, though the brief did not find Google stating that directly.

Hydrogen is not a stripped-down Carbon and not a replacement for it. In its Hydrogen announcement, Google describes the frontier models, Carbon and Argon, as strongest at deep, long-horizon reasoning, and Hydrogen as built for fast execution across varied agentic workflows. Google also called Carbon its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities.

Some pre-launch rumor coverage called Hydrogen "Gemini 4 Flash." One tracker that read the relevant pages found that none of them use that name, so this article uses Hydrogen. Both models are hosted and cloud-only. Carbon has no published weights or parameter count, and Google has not announced downloadable weights for Hydrogen.

What the benchmarks show

The most widely cited comparison is the Artificial Analysis Intelligence Index, a composite score that Artificial Analysis publishes. Google says Hydrogen edges past Argon on it at its default high thinking setting, 51 against 50, and reaches 54 at the top xhigh setting. A secondary source puts Carbon at high thinking at 60, so Carbon leads on this index.

The table below, compiled by the API reseller AIHubMix from Artificial Analysis data, shows how score, cost, time and responsiveness trade off across settings.

Artificial Analysis Intelligence Index results by model setting, November 2026 (as reported by AIHubMix; cost and time are per index task)
Setting Leaderboard rank AA Index Cost per task Time per task First answer token
Carbon high 4 60 $5.48 28.1 min 63.37 s
Hydrogen xhigh 16 54 $3.16 13.5 min 37.77 s
Hydrogen high (default) 21 51 $1.87 8.3 min 18.89 s
Hydrogen medium 39 41 $0.80 3.7 min 11.80 s
Hydrogen low 60 27 $0.29 1.3 min 5.71 s

The pattern is one of diminishing returns. Moving from Hydrogen medium to high adds 10 index points for about $1.07 more per task, while the step from high to xhigh adds 3 points for about $1.29 more. By the same source's comparison, Carbon at high is six points ahead of Hydrogen at xhigh but costs $2.32 more per task, 73.4% more, and takes 28.1 minutes against 13.5.

Speed

Speed is where the gap is clearest. The aggregator TokenHarbor, citing Artificial Analysis figures as of November 13, lists Hydrogen at roughly 190 output tokens per second against about 73 for Carbon and 63 for Argon. Other sources report 188, 198 or 236 for Hydrogen and 72 or 76 for Carbon, depending on thinking level and date, but every one shows Hydrogen at least about 2.5 times faster.

Other reported results

The reseller OrcaRouter reports 71.6% for Hydrogen and 77.4% for Carbon on Agents' Last Exam. On Vending-Bench 2, a test scored by final account balance, it lists $14,549 for Hydrogen and $16,484 for Carbon. MarkTechPost reports Hydrogen at 57% on Terminal-Bench 4.1, against 18.2% for the earlier Gemini 3.8 Flash, and 58.8% on HLE-Diamond, without giving Carbon's figures in the same source. The direction matches the index results: Carbon scores higher where both are reported.

Thinking settings

Secondary sources relaying Google's thinking documentation say Hydrogen defaults to high and accepts low, medium, high and xhigh. Carbon defaults to medium and stops at high. The reseller Apiyi describes Carbon as having three tiers, and neither it nor the brief verified the primary documentation.

More thinking does not always mean a better result. Reports on Artificial Analysis data show Carbon at low scoring 54 against 58 at medium, at a higher cost per task ($5.22 against $4.68). Settings are a variable to test, not a dial that only moves one way.

Price: a gap that widens on December 1

On the Gemini API, Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens, flat across its 1 million-token context window, according to Google. Cached input is $0.075 per million. No introductory period has been reported for it.

Carbon is available through the API in AI Studio and Gemini for business to paying developers and enterprises, at the same introductory rate as Argon: $2 input and $10 output. The analysis site Neomanex, quoting Google's footnote, reports that $4 input and $20 output will apply after November 30, 2026. Another outlet says Google's pricing page confirms those rates from December 1. The brief did not open the pricing page, so the increase is probable but worth verifying. The benchmark site benchlm.ai lists Carbon's cached-input rate at $0.10 now and $0.20 from December 1.

Jaroha's arithmetic from those list prices puts Carbon at about 1.33 times Hydrogen's per-token cost now and about 2.67 times from December 1. One outlet reaches the same figures. Free access differs too: Google's pricing page reportedly says Carbon and Argon are not available on the free tier, while Hydrogen is, under Google's usual limits.

Which tasks suit each model

Where Hydrogen fits

Google positions Hydrogen for rapid execution, and the third-party guidance in the brief points the same way: short code-and-test loops, high-volume work, latency-sensitive jobs and free-tier trials. The speed and per-task cost figures above explain why. A task that needs many quick iterations benefits from the 13.5-minute or 8.3-minute index runs more than from a few extra score points.

Where Carbon fits

Carbon is the intended tool for planning, hard debugging, architecture decisions, scientific analysis and security work, based on Google's positioning and third-party tests. A secondary source attributes to Google's availability post the view that Argon remains the better pick for most software engineering and knowledge work, with Carbon reserved for tasks that push a model's limits. The brief could not verify that wording against the original, but it is consistent with Argon sitting in the middle of the lineup.

Third-party analysis also describes Carbon as dual-use. The ability that lets a model find and patch software vulnerabilities can also let it find and exploit them. Google has kept a Fairwind program for cyber defenders going, and a free six-month program for scientists has also been reported.

Using them together

Google's own suggestion is to combine them. Carbon acts as coordinator, while Hydrogen powers worker and verifier agents that build and test changes in parallel. Google reports that on Terminal-Bench 4.1, a Carbon-plus-Hydrogen pairing reached near-Carbon task success while using 24% less quota and finishing 19% faster than Carbon alone.

That result is the only quantified evidence for the planner-and-executor idea, and it comes from the vendor. Apiyi calls the same test Terminal-Bench 4.0 rather than 4.1, a discrepancy the brief could not resolve. Apiyi also says Google does not put the two models side by side in a single per-benchmark table, so head-to-head comparisons come from third parties, several of which sell API access and gain from model traffic.

Who can use each model

Hydrogen has the broader reach. It is available on free and paid Gemini API and AI Studio tiers, and paid Google AI Pro and Ultra subscribers get it in Antigravity. It also appears in Search. In the Gemini app, free users get it through automatic model selection "when appropriate," and Google says rollout may vary. Choosing Hydrogen manually requires Google AI Plus at $4.99 a month, according to Fello AI, a secondary source.

Orchestrator, a new Antigravity picker option that pairs Carbon and Hydrogen, is rolling out only to AI Ultra and Gemini for business. Pro users get Hydrogen but not Orchestrator for now.

Carbon is narrower. Google says everyday users can try it in the Gemini app and in AI Mode in Search if they subscribe to Google AI Ultra. Enterprises can reach it in Gemini for business through the API or a model drop-down, in Google Antigravity, or in CodeMender. Google AI Pro subscribers get Argon instead, and nothing in the brief says that will change.

On cloud platforms, Hydrogen is documented on Google Cloud's enterprise agent platform. Carbon's launch materials name the Gemini API in AI Studio and Gemini for business without mentioning Vertex AI, so its status on Google's enterprise cloud platform is unconfirmed.

Limits of the evidence

No peer-reviewed or independent academic study of either model turned up. Performance evidence is Google's own claims or numbers from benchmark services and commercial blogs. Many articles repeat the same Google posts and model cards, so a large count of articles is not independent confirmation.

Several figures also conflict. Carbon scored 59 on the index in launch-day figures and 60 in later comparisons, possibly from a re-run or different settings, which is unconfirmed. Carbon's cost per task appears as $5.48 at AIHubMix and OrcaRouter but $7.76 at Fello AI and NeuralTrust, and the sources do not explain the difference. The $5.48 figure in the table should therefore be checked against Artificial Analysis directly.

Hydrogen's one-point lead over Argon (51 to 50) on a single index sits in a range where settings, run noise or index version could matter, and nobody reports confidence intervals. A composite index also says little about how a model handles one specific task. Cost per index task is likewise not cost per real job.

A division of labor with a pricing clock

The two models are best read as a pair rather than a ranking. Carbon posts the higher scores and costs more in money and time. Hydrogen trades some of that headroom for speed and a lower rate, and it reaches far more users. For most readers outside engineering teams, the practical question is less about benchmarks than about which plan unlocks which model.

Three things are worth watching. The first is whether Carbon's API price moves to $4 and $20 on December 1, which would roughly double the cost gap. The second is whether independent testers publish repeated, task-level comparisons that show how much Carbon's lead matters on everyday work. The third is whether the pairing result from Google holds up when others run it.

Frequently asked questions

When was Gemini 4 Hydrogen released?

November 12, 2026. Google Cloud's model documentation lists that date, and Artificial Analysis and a tracker citing Google's launch post agree. Coverage dated November 13 reflects when articles were published, not the release.

Is Gemini 4 Hydrogen just a cheaper Carbon?

Not exactly. Google's model card says Hydrogen is adapted from Carbon, but Google presents it as an execution model for fast agentic work. Carbon posts higher scores on the benchmarks reported so far, while Hydrogen is faster and cheaper.

How much do Gemini 4 Carbon and Hydrogen cost?

On the Gemini API, Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens. Carbon costs $2 and $10 through November 30, 2026, and is reported to rise to $4 and $20 from December 1. The pricing page itself was not verified for this article.

Can free users access Carbon or Hydrogen?

Hydrogen is on the Gemini API free tier under Google's usual limits, and free Gemini app users may get it through automatic model selection. Carbon is reportedly not on the free tier, and in the Gemini app it is limited to Google AI Ultra subscribers.

Which Gemini 4 model should developers start with?

Google's guidance and third-party tests suggest Hydrogen for short code-and-test loops, high volume and latency-sensitive work. Carbon is intended for planning, hard debugging, architecture and security tasks, with Argon as the middle option. Much of that advice comes from vendors and resellers, so testing on real workloads is the safer check.

Sources (18)

  1. Gemini 4 Hydrogen | Gemini Enterprise Agent Platform | Google Cloud Documentation , Google Cloud
  2. Gemini 4 Hydrogen Model card , Google DeepMind
  3. Introducing Gemini 4 Hydrogen for Agentic Workflows , Google (2026-11-12)
  4. Gemini 4 Carbon is now available , Google (2026-11-10)
  5. Gemini 4 Carbon: our next frontier model , Google (2026-10-28)
  6. Gemini 4 Hydrogen Models - Intelligence, Performance & Price Comparison , Artificial Analysis
  7. Google Releases Gemini 4 Hydrogen: 2.5x Faster Than Gemini 3.8 Flash , MarkTechPost (2026-11-13)
  8. Gemini 4 Hydrogen Launch: Google, Nov 12, 2026 , Kingy AI (2026-11-13)
  9. Google makes Gemini 4 Carbon available at Argon's introductory price , Runtime Wire (2026-11-10)
  10. Gemini 4 Carbon Is GA on the Paid Gemini API: $2/$10 Until 30 November, Then $4/$20 , Neomanex
  11. Gemini API Pricing (November 2026): Rates & Price Increases , benchlm.ai
  12. Gemini 4 Hydrogen: Price, Benchmarks and Who Gets It , Fello AI
  13. Gemini 4 Carbon vs Hydrogen vs Argon: Which Model for Which Job? , AIHubMix
  14. Gemini 4 Hydrogen vs Gemini 4 Carbon: Two Price Calendars , OrcaRouter
  15. Gemini 4 Hydrogen vs Argon vs Carbon: Which Model Should You Use? , TokenHarbor (2026-11-13)
  16. Gemini 4 Hydrogen vs Gemini 4 Carbon: How to Choose for Everyday Use , Apiyi.com Blog
  17. Gemini 4 Carbon: Google's Top Model Is Cloud Only, No Weights , modelfit.io
  18. Gemini 4 Hydrogen is Google's new workhorse for AI agents, and what European teams should check first , i-SCOOP

Share this article

Back to list