Skip to content
, ,

Gemini 4 Hydrogen vs Gemini 4 Carbon (Nov 2026): Which Should You Use?

11–16 minutes
Winking woman in a blue-violet shirt on a park bench holding a Gemini logo card, hook Hydrogen or Carbon, which Gemini

Google shipped two Gemini 4 models in the same week. Gemini 4 Carbon opened to paying developers and AI Ultra users on 10 November 2026, and Gemini 4 Hydrogen followed on 12 November as the new fast, cheap default.

They share the same 1M-token window, but they cost, think and feel very different. This guide helps you pick the right one for each job, with real prices, worked cost maths and copy-paste API code.

If you want the full deep dive on one model, read my Gemini 4 Hydrogen guide or the Gemini 4 Carbon benchmarks and access post. This page is the side-by-side decision piece.

TL;DR

  • Default pick: Hydrogen (gemini-4-hydrogen) at $1.50 / $7.50 per 1M tokens. It is fast, has a free API tier and handles most coding, chat and agent work.
  • Hard problems: Carbon (gemini-4-carbon) scores 60 on the Artificial Analysis Intelligence Index vs Hydrogen’s 51 to 54, and leads MathArena.
  • Price change: Carbon is $2 / $10 until 30 November 2026, then $4 / $20 from 1 December. Hydrogen’s price does not change.
  • Speed: Artificial Analysis measures Hydrogen at about 170 tokens/s vs about 62 for Carbon at high effort.
  • Access: Hydrogen is in the free Gemini app (auto mode), Search and every API tier. Carbon needs AI Ultra or the paid API tier.

Quick Answer: Which One Should You Use?

Quick answer: Start every project on Hydrogen. Move a task to Carbon only when Hydrogen fails it, or when the task is long, high-stakes reasoning such as hard maths, security review or a multi-hour agent run.

Google itself frames the pair this way. It calls Hydrogen “our daily driver for agentic workflows”, while Carbon and Argon are the frontier models for deep, long-horizon reasoning.

  • Pick Hydrogen for: chat, quick coding edits, tool calls, data extraction, support bots, high-volume pipelines and anything a user waits on.
  • Pick Carbon for: hard maths, tricky debugging across a big codebase, security audits, legal or finance analysis where one mistake is expensive, and planning steps in long agent runs.
  • Pick both: let Carbon plan and Hydrogen execute. That is exactly what Google’s Antigravity Orchestrator does.

Specs Side by Side

Both models come from Google’s Gemini Enterprise Agent Platform model pages and the Gemini API docs. The biggest differences are thinking levels, default effort and free-tier access.

Spec Gemini 4 Hydrogen Gemini 4 Carbon
API model ID gemini-4-hydrogen gemini-4-carbon
Released 12 Nov 2026 GA 10 Nov 2026 (cyber preview from 28 Oct)
Role Fast daily driver, replaces 3.8 Flash Frontier reasoning and long-horizon agents
Context window 1,048,576 tokens, shared by input, thinking and output 1,048,576 tokens, shared the same way
Max output per segment 65,536 tokens 65,536 streaming, 32,768 non-streaming (Cloud docs)
Long decode Up to about 1M tokens with background mode Up to about 1M tokens with background mode
Thinking levels low, medium, high (default), xhigh low, medium (default), high
Inputs Text, image, audio, video, PDF, Office files Text, image, audio, video, PDF, Office files
Caching Implicit only, 4,096-token minimum Implicit only, 4,096-token minimum
Sampling (temperature etc.) Fixed; custom values ignored Fixed; custom values ignored
Gemini API free tier Yes, with rate limits No, paid tier only
Cloud launch stage “Short-term” GA, 45-day retirement notice GA (cyber variant allowlist-only)

Note: Hydrogen is a “short-term” GA release on Google Cloud. Google can retire that exact model ID with 45 days’ notice, so keep the model name in a config value, not hard-coded.

Pricing, Including the 1 December Change

Hydrogen has one flat price across the full 1M window. Carbon launched at an introductory rate that doubles on 1 December 2026. Thinking tokens bill as output on both.

Per 1M tokens Hydrogen Carbon until 30 Nov Carbon from 1 Dec
Input $1.50 $2.00 $4.00
Output (incl. thinking) $7.50 $10.00 $20.00
Cached input (implicit hit) $0.075 $0.10 $0.20
Batch / Flex $0.75 / $3.75 $1.00 / $5.00 Check the pricing page
Free API tier Yes No No

Google confirmed the $2 / $10 to $4 / $20 step for Carbon, and 9to5Google reports the intro rate runs until 30 November. Batch and cached rates for Carbon come from LiteLLM’s launch notes, which match Google’s 95% cache discount.

Warning: if you prototype on Carbon in November, your bill for the same workload doubles in December. Budget at the $4 / $20 rate from day one.

For every other Gemini model, free-tier limits and how billing works, see my Gemini API pricing and free tier guide.

Speed and Benchmarks

Speed: Hydrogen wins clearly

Google says Hydrogen finishes tasks 2.5x faster than Gemini 3.8 Flash at its default high setting. Early partners saw 2x to 2.8x faster runs with 12% to 23% fewer tokens. Independent numbers from Artificial Analysis (AA) show how big the gap to Carbon is.

Measure (Artificial Analysis) Hydrogen (high) Carbon (high) Carbon (medium)
Output speed ~170 tokens/s ~62 tokens/s ~59 tokens/s
Time to first answer token ~16 s ~72 s ~22 s
Average time per AA task 8.3 min 31.3 min 28.1 min

That 72-second wait at Carbon high is the thinking phase. It is fine for a background job, but painful in a chat window.

Benchmarks: Carbon is smarter, but not everywhere

Carbon leads on broad intelligence and maths. On some professional agent tests, though, Hydrogen matches it. Sources are listed in each row.

Benchmark Hydrogen Carbon Source
AA Intelligence Index 27 low, 41 medium, 51 high, 54 xhigh 54 low, 58 medium, 60 high Artificial Analysis
Terminal-Bench 4.0 54% (high and xhigh) 65.2% (high) Artificial Analysis
MathArena expected performance Not listed yet 85.4% (high), ranked #1 MathArena
Vals Tax Agent Bench 52.3% 51.78% Google citing Vals / Vals
Vals Legal Research Bench 55.8% 55.77% Google citing Vals / Vals
HLE Diamond 58.8% About 69% Google / AA

Notice that Hydrogen at xhigh (54) only matches Carbon at low (54). So for the hardest reasoning, Carbon’s floor is Hydrogen’s ceiling.

On accuracy, Google says Hydrogen makes more than 5x fewer fabricated claims than 3.8 Flash. For Carbon, The Next Web noted its hallucination rate on one AA knowledge test rose at launch, so check facts in research answers either way.

Which Plan or App Gets Which Model

Consumer access is where the two models differ most. Hydrogen is everywhere, while Carbon sits behind the top plan or a paid API key.

Where Hydrogen Carbon
Gemini app, free Yes, via automatic model selection No
Gemini app, paid subscribers Yes, in the model dropdown AI Ultra only (“Most complex work”)
Google Search All users worldwide AI Mode for AI Ultra users
Google Antigravity AI Pro and Ultra AI Ultra
Gemini in Google Sheets AI Pro, Ultra and eligible business plans Not announced
Gemini API / AI Studio All tiers, including free Paid tier only
Android Studio Yes Yes (paid API)
Google Cloud Agent Platform GA GA

VentureBeat lists AI Ultra at $99.99 a month (Ultra 5x) and $199.99 (Ultra 20x). AI Pro users keep Gemini 4 Argon alongside Hydrogen. My Google AI Pro vs Ultra guide breaks down the limits per plan.

Tip: in Antigravity, the Orchestrator lets Carbon plan while Hydrogen runs worker and verifier agents. Google says this got near Carbon’s Terminal-Bench 4.0 success rate while using about 24% less quota and finishing about 19% faster.

Thinking Levels Compared

Both models use thinking_level instead of a token budget. The names overlap, but the defaults do not.

  • Hydrogen: low, medium, high (default) and xhigh. AA shows xhigh costs about 70% more per task than high for only 3 extra index points.
  • Carbon: low, medium (default) and high. Medium is the sweet spot: 58 on the AA index with a 22-second first token, vs 60 and 72 seconds at high.
  • Both reject: thinking_budget and MINIMAL, which return an HTTP 400 error.

A useful mental model: Hydrogen high is the “normal” setting for agents, and Carbon medium is the “careful expert” setting. Reach for Carbon high only on problems where an extra two index points matter more than a minute of waiting.

Cost per Task: Worked Examples

These are worked calculations from the official per-token prices, not test runs. Real thinking-token counts vary, and Carbon usually thinks longer than Hydrogen on the same prompt, which widens the gap further.

Example 1: a support chatbot reply

Assume 3,000 input tokens and 800 output tokens (including thinking) per reply, for 10,000 replies a month.

  • Hydrogen: (3,000 x $1.50 + 800 x $7.50) / 1M = $0.0105 per reply, or $105 a month.
  • Carbon in November: (3,000 x $2 + 800 x $10) / 1M = $0.014, or $140 a month.
  • Carbon from December: (3,000 x $4 + 800 x $20) / 1M = $0.028, or $280 a month.

Example 2: a coding agent session

Assume 400,000 input tokens across all turns, of which 300,000 are implicit cache hits, plus 40,000 output tokens.

  • Hydrogen: 100K x $1.50 + 300K x $0.075 + 40K x $7.50 = $0.15 + $0.0225 + $0.30 = about $0.47.
  • Carbon in November: $0.20 + $0.03 + $0.40 = about $0.63.
  • Carbon from December: $0.40 + $0.06 + $0.80 = about $1.26.

Example 3: one big document review

Assume an 800,000-token input (a large contract set or codebase) and a 10,000-token report, sent once with no cache hit.

  • Hydrogen: $1.20 + $0.075 = about $1.28.
  • Carbon in November: $1.60 + $0.10 = $1.70.
  • Carbon from December: $3.20 + $0.20 = $3.40.

What independent testers measured

Artificial Analysis runs the same task suite on every model, so its cost per task includes real thinking overhead. Current figures on its Google page:

Model and level Avg cost per AA task
Hydrogen low / medium / high / xhigh $0.29 / $0.80 / $1.87 / $3.16
Carbon low / medium / high $5.22 / $4.68 / $5.48

These Carbon figures reflect the $2 / $10 intro price. At $4 / $20 the same runs would cost roughly twice as much, so Hydrogen high works out about 5x cheaper per task than Carbon medium after 1 December.

How to Switch Between Them in the API

Switching is a one-line model change, plus a thinking level that exists on the target model. Here is the safe way to set it up.

  1. Get a key. Create a Gemini API key in Google AI Studio. For Carbon, turn on billing so the key is on the paid tier.
  2. Install the SDK. Run pip install -U google-genai and set GEMINI_API_KEY in your environment.
  3. Put the model in config. Read the model ID and thinking level from environment variables, so you can switch without a code change.
  4. Remove old parameters. Delete temperature, top_p, top_k, thinking_budget and any explicit cached_content fields.
  5. Map thinking levels. Use xhigh only with Hydrogen. Carbon tops out at high.
  6. Compare on your own evals. Run the same 20 to 50 real prompts through both models and log quality, latency and the usage token counts before you commit.

Python: Interactions API

Google’s Hydrogen quickstart uses the new Interactions API. This version reads the model from config and picks a valid thinking level.

import os
from google import genai

MODEL = os.environ.get("GEMINI_MODEL", "gemini-4-hydrogen")
LEVEL = os.environ.get("GEMINI_THINKING", "high")

# Carbon has no xhigh level
if MODEL == "gemini-4-carbon" and LEVEL == "xhigh":
    LEVEL = "high"

client = genai.Client()  # reads GEMINI_API_KEY

interaction = client.interactions.create(
    model=MODEL,
    input="Find the race condition in this retry loop and fix it.",
    generation_config={"thinking_level": LEVEL},
)

print(interaction.output_text)
print(interaction.usage)

Python: a simple Hydrogen-first router

This pattern sends everything to Hydrogen and escalates to Carbon only for tasks you flag as hard.

def ask(prompt, hard=False):
    model = "gemini-4-carbon" if hard else "gemini-4-hydrogen"
    level = "medium" if hard else "high"
    result = client.interactions.create(
        model=model,
        input=prompt,
        generation_config={"thinking_level": level},
    )
    return result.output_text

print(ask("Summarise this ticket in two lines: ..."))
print(ask("Audit this auth flow for privilege escalation: ...", hard=True))

REST with curl

export GEMINI_MODEL="gemini-4-carbon"

curl "https://generativelanguage.googleapis.com/v1beta/interactions" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H 'Content-Type: application/json' \
  -X POST \
  -d '{
    "model": "'"$GEMINI_MODEL"'",
    "input": "Explain this stack trace and suggest a fix.",
    "generation_config": {"thinking_level": "medium"}
  }'

Decision Table by Use Case

Use case Pick Suggested level Why
Chatbot, support, FAQ answers Hydrogen medium Fast first token, lowest cost
Classification, extraction, tagging Hydrogen low $0.29 per AA task at low; Batch halves it again
Everyday coding and quick fixes Hydrogen high About 170 tokens/s and solid Terminal-Bench
Agent worker or verifier steps Hydrogen high Google’s own Orchestrator uses it this way
Long-horizon agent planning Carbon medium Higher success on long tasks
Hard maths and proofs Carbon high #1 on MathArena at 85.4%
Security review of code Carbon medium or high Built for cyber work
Legal or tax research agents Try Hydrogen first high Vals scores are nearly tied
Free experiments, students Hydrogen any Only one with a free API tier
Big document analysis (500K+ tokens) Hydrogen, then Carbon on misses high Same 1M window, less than half the cost after 1 Dec

Troubleshooting: Common Problems and Fixes

  • Carbon call fails on a free key: Carbon is paid tier only. Enable billing on the AI Studio project or switch the model to Hydrogen.
  • 400 error on thinking_budget or MINIMAL: neither model supports them. Use thinking_level with a valid value.
  • Error when sending xhigh to Carbon: Carbon only lists low, medium and high. Map xhigh to high in your code.
  • 400 error on temperature: Hydrogen ignores in-range sampling values and rejects out-of-range ones. Remove temperature, top_p and top_k.
  • Explicit cache calls fail: neither model supports caches.create at launch. Rely on implicit caching by keeping long shared prefixes at the start of the prompt.
  • Output stops at 65,536 tokens: that is the per-segment cap. Use background mode with a higher max_output_tokens for longer outputs.
  • Carbon feels frozen in chat: at high effort AA measured about 72 seconds before the first answer token. Use medium, stream the response, or move the job to the background.
  • Carbon missing from the Gemini app: it is only for AI Ultra subscribers. Pro users see Argon and Hydrogen.
  • Bill jumped in December: Carbon’s intro price ended on 30 November. Re-run your cost maths at $4 / $20.

For quota and rate-limit errors such as 429 and 503, see my Gemini API errors fix guide.

Frequently Asked Questions

Is Gemini 4 Hydrogen better than Gemini 4 Carbon?

No, Carbon is the stronger reasoner, scoring 60 on the Artificial Analysis Intelligence Index vs 51 for Hydrogen at default. Hydrogen is faster and much cheaper, which makes it the better choice for most everyday tasks.

What are the API model IDs?

Use gemini-4-hydrogen for Hydrogen and gemini-4-carbon for Carbon. A separate gemini-4-carbon-cyber-permissive variant exists for vetted security teams in Google’s Fairwind Program.

When does Gemini 4 Carbon’s price go up?

The $2 / $10 introductory price runs until 30 November 2026. From 1 December it rises to $4 per 1M input tokens and $20 per 1M output tokens.

Can I use Gemini 4 Carbon for free?

Not in the API, where Carbon is paid tier only. In Google’s apps it is limited to AI Ultra subscribers, while Hydrogen is free in the Gemini app and has a free API tier.

Do both models have the same context window?

Yes, both have a 1,048,576-token window shared by input, thinking and output. Each foreground response segment is capped at 65,536 tokens, with longer outputs available through background mode.

Does Hydrogen replace Gemini 3.8 Flash?

Yes, Hydrogen is the new fast tier that takes over from 3.8 Flash. To migrate, change the model ID, drop sampling parameters and explicit caching, and note the default thinking level moves from medium to high.

Can I use both models together?

Yes. A common pattern is Carbon for planning and Hydrogen for execution, which is how Google’s Antigravity Orchestrator works, and you can copy it with the simple router shown above.

Sources

Free weekly email

Get new AI guides every week

One short weekly email with our newest local AI, ComfyUI and AI tool guides. No spam, no pop-ups, just the good stuff.

Free. Unsubscribe anytime in one click. We only use your email to send this newsletter. See our privacy policy.

Winking woman in a blue-violet shirt on a park bench holding a Gemini logo card, hook Hydrogen or Carbon, which Gemini