# Gemini 4 Hydrogen vs Carbon: Coding and Token Cost

> Compare Hydrogen and Carbon for coding and agent tasks, with reasoning settings, task latency, token use and the limits of published benchmarks.

- Canonical HTML: https://omniakey.com/blog/gemini-4-hydrogen-vs-carbon
- Markdown: https://omniakey.com/blog/gemini-4-hydrogen-vs-carbon.md
- Published: 2026-11-13 (updated 2026-11-13)
- Author: Kiki, COO at OmniaKey
- Language: en
- Category: Comparison
- Reading time: 8 min
- All posts: https://omniakey.com/blog/llms.txt

---

Gemini 4 Neon was released on November 18, 2026. The [Gemini 4 Neon status guide](https://omniakey.com/blog/gemini-4-neon) has the model details, and the [Neon vs Carbon comparison](https://omniakey.com/blog/gemini-4-neon-vs-carbon) covers coding, speed and cost.

**Start by trying [Gemini 4 Hydrogen](https://omniakey.com/gemini-4-hydrogen) for responsive edits; evaluate Carbon for harder repository work, where Artificial Analysis's two coding tests scored it higher.** Artificial Analysis tested both at high effort.

## Hydrogen vs Carbon at a glance

Check the defaults first, then the access and speed that affect the bill.

|Item|Hydrogen|Carbon|
|-|-|-|
|Google's positioning|Everyday coding, tool agents|Harder debugging, architecture|
|Default thinking|high|medium|
|Highest thinking|xhigh|high|
|Free tier|Yes, rate-limited|Paid API only|
|Context window|1,048,576|1,048,576|
|Output speed (AA)|179 tokens/s|72 tokens/s|

The positioning row is Google's own description, not a measured result. The measured differences here are Carbon's higher scores on two coding tests, and Hydrogen's lower Standard unit rate and faster generation throughput.

Both are text-output models with a 1,048,576-token context window, and Google lists the same tool set for each ([Hydrogen model page](https://ai.google.dev/gemini-api/docs/models/gemini-4-hydrogen), [Carbon model page](https://ai.google.dev/gemini-api/docs/models/gemini-4-carbon)).

## What independent coding results show

Artificial Analysis's model pages for [Hydrogen (high)](https://artificialanalysis.ai/models/gemini-4-hydrogen-high) and [Carbon](https://artificialanalysis.ai/models/gemini-4-carbon), snapshot November 13, 2026, use the same Intelligence Index v5.0.1 across nine evaluations, with a grading update on November 3. At high effort, Carbon scores 60 and Hydrogen 51 on that broad index, which is not coding-only.

|Test|Hydrogen high|Carbon high|
|-|-:|-:|
|Terminal-Bench 4.0|54.0%|65.2%|
|FrontierCode-AA 2.0|36.3%|40.6%|

[Terminal-Bench 4.0](https://artificialanalysis.ai/evaluations/terminalbench-4-0) runs 66 tasks three times each on the mini-swe-agent harness, scores pass\@1 with a 500-step cap, and requires every verifier test to pass. It rewards an agent that can finish a terminal task end to end. [FrontierCode-AA 2.0](https://artificialanalysis.ai/evaluations/frontiercode-2-0) runs 132 tasks with three attempts on the Stirrup harness and a 500-turn cap. It uses pass/fail checks and a rubric judge (GPT-6 Luna at medium), and reports a weighted quality score, not a binary pass rate.

We have not run a matched two-model test; every score in this section comes from Artificial Analysis. Hydrogen's xhigh setting scored 53.5% on Terminal-Bench 4.0 and 39.0% on FrontierCode-AA 2.0, with an index of 54. Higher effort did not improve every displayed score: Terminal-Bench was slightly lower and FrontierCode higher.

Artificial Analysis also measured 179 output tokens per second for Hydrogen high and 72 for Carbon high, sampling endpoints every six hours over the trailing five days ([Hydrogen](https://artificialanalysis.ai/models/gemini-4-hydrogen-high), [Carbon](https://artificialanalysis.ai/models/gemini-4-carbon)). That is generation throughput, not the wall-clock time of a repository job; tool calls, retries and queueing add to the total.

## Default thinking levels are not equal effort

Hydrogen defaults to high and accepts low, medium, high and xhigh; Carbon defaults to medium and stops at high ([Google's thinking docs](https://ai.google.dev/gemini-api/docs/thinking)). Neither accepts minimal. Google publishes no fixed token budget per level, so a default call, or the same label, is not a controlled equal-compute comparison.

Set `thinking_level` explicitly when you compare them, and treat the level as part of the task setup. Thinking tokens are billed as output, so a larger reasoning budget can cost more than the lower per-token rate saves. xhigh suits hard multi-step coding and agent problems; Google warns that it can overthink simple prompts ([Hydrogen what's new](https://ai.google.dev/gemini-api/docs/whats-new-gemini-4-hydrogen)).

## Output limits for coding agents

In Google's Interactions API, the foreground non-streaming limit is 65,536 output tokens for Hydrogen and 32,768 for Carbon ([Google's thinking docs](https://ai.google.dev/gemini-api/docs/thinking)). Streaming is 65,536 for both ([Hydrogen model page](https://ai.google.dev/gemini-api/docs/models/gemini-4-hydrogen), [Carbon model page](https://ai.google.dev/gemini-api/docs/models/gemini-4-carbon)).

Above 65,536, both support long decode with `background` execution or client-side continuation, up to 1,048,576 minus the input ([Hydrogen what's new](https://ai.google.dev/gemini-api/docs/whats-new-gemini-4-hydrogen), [Carbon what's new](https://ai.google.dev/gemini-api/docs/whats-new-gemini-4-carbon)).

The context window covers input, thinking and output together, so a large repository prompt leaves less room for output. If input plus `max_output_tokens` exceeds the window, the API reduces the output cap to the remaining context.

`max_output_tokens` caps thinking plus output, so a hard cap can end a reasoning turn as `incomplete` while the generated thinking tokens are still billed. An agent that sends non-streaming requests should plan around Carbon's smaller limit.

Hydrogen's [setup guide](https://omniakey.com/blog/gemini-4-hydrogen) covers access, thinking levels and output limits.

## Official price and one worked example

Google's Standard rates on November 13, 2026 are below; Carbon's current rate runs through November 30, 2026 ([pricing page](https://ai.google.dev/gemini-api/docs/pricing)).

|Standard rate|Hydrogen|Carbon|
|-|-:|-:|
|Input / 1M|$1.50|$2.00|
|Cached input / 1M|$0.075|$0.10|
|Output / 1M|$7.50|$10.00|

Output tokens include thinking. From December 1, 2026, Carbon's Standard rate becomes $4.00 input, $0.20 cached input and $20.00 output per million tokens. Hydrogen has one flat rate with no end date.

Implicit caching is eligible on prompts of at least 4,096 tokens and discounts cached input by 95%; a hit requires a matching, reused prompt prefix, so length alone does not guarantee one ([pricing page](https://ai.google.dev/gemini-api/docs/pricing)).

For 100,000 uncached input tokens and 20,000 total output tokens including thinking, Hydrogen costs 0.1 × $1.50 + 0.02 × $7.50 = $0.30. Carbon costs 0.1 × $2.00 + 0.02 × $10.00 = $0.40 through November 30, and 0.1 × $4.00 + 0.02 × $20.00 = $0.80 from December 1. Retries, tool rounds, cache hits and extra reasoning tokens move the bill, and the same visible answer length does not mean the same token count.

Hydrogen's free tier is rate-limited, and Google marks free-tier content as used to improve its products; the paid tier is marked no ([Hydrogen model page](https://ai.google.dev/gemini-api/docs/models/gemini-4-hydrogen)). Carbon has no free tier, so any Carbon test starts on a paid key ([Carbon model page](https://ai.google.dev/gemini-api/docs/models/gemini-4-carbon)).

## How to choose, migrate and test

Google positions Carbon for harder repository work and Hydrogen for faster everyday coding; treat that as vendor guidance, not a guarantee. Its pairing advice is to let Carbon plan interfaces or clear roadblocks, then hand edits, parallel subagent tasks and test runs to Hydrogen ([Google's announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/)). The measured case is narrower: Carbon scored higher on the two coding tests, while Hydrogen has the lower Standard unit rate and faster measured generation.

- **Try Hydrogen first** if most of your work is short edit-test loops, you want a rate-limited free tier to try prompts, or response speed and unit price drive the bill.
- **Evaluate Carbon** if a task keeps failing at planning or verification and the higher unit price and slower generation are worth the extra coding score.
- **Migrate within Google's API** by changing the model ID and setting `thinking_level` explicitly. Log correctness, tokens, tool rounds and wall time on the same tasks, then escalate only the tasks that fail.
- **On OmniaKey**, Hydrogen is listed at its model page; Carbon runs through Google's paid Gemini API. OmniaKey's listing, top-up options and prices are separate from Google's API terms and prices.

Within the same Google API and client, switching between these two models or from Argon is a request-format change. Set the model ID (`gemini-4-hydrogen` or `gemini-4-carbon`) and `thinking_level` ([Google's thinking docs](https://ai.google.dev/gemini-api/docs/thinking)). Both ignore `temperature`, `top_p` and `top_k` if sent, and neither supports explicit context caching, so remove those fields. Retry or escalate according to your team's verifier; compare accepted output, tokens and wall time before making the switch permanent. Moving to another provider or a different request format can require more changes.

Teams already on Argon can read the [Carbon vs Argon comparison](https://omniakey.com/blog/gemini-4-carbon-vs-argon). The earlier [Argon coding test](https://omniakey.com/blog/gemini-4-argon-coding-test) compared Argon with Claude Opus 5.5 on eight tasks; Hydrogen was not part of that test.

## Frequently asked questions

### Which model is better for coding, Hydrogen or Carbon?

At high effort, Artificial Analysis scored Carbon higher on both coding tests in its November 13, 2026 snapshot: 65.2% against 54.0% on Terminal-Bench 4.0, and 40.6% against 36.3% on FrontierCode-AA 2.0 ([Hydrogen](https://artificialanalysis.ai/models/gemini-4-hydrogen-high), [Carbon](https://artificialanalysis.ai/models/gemini-4-carbon)). Try Hydrogen first when speed and price matter, and evaluate Carbon when a task keeps failing planning or verification.

### Do Hydrogen and Carbon use the same thinking level by default?

No. Hydrogen defaults to high and supports low, medium, high and xhigh; Carbon defaults to medium and stops at high ([Google's thinking docs](https://ai.google.dev/gemini-api/docs/thinking)). Google publishes no fixed token budget per level, so set `thinking_level` explicitly before comparing cost or quality.

### What does one request cost?

At Google's Standard rates on November 13, 2026, 100,000 uncached input tokens plus 20,000 output tokens including thinking costs $0.30 on Hydrogen, $0.40 on Carbon through November 30, and $0.80 on Carbon from December 1 ([pricing page](https://ai.google.dev/gemini-api/docs/pricing)). Retries, tools and cache hits change the total.

### Is 179 versus 72 output tokens per second a task-time comparison?

No. Those Artificial Analysis figures measure generation throughput, sampled every six hours over the trailing five days ([Hydrogen](https://artificialanalysis.ai/models/gemini-4-hydrogen-high), [Carbon](https://artificialanalysis.ai/models/gemini-4-carbon)). A repository job also includes tool calls, retries and queueing, so measure wall time on your own tasks.

### Can I call Carbon through OmniaKey?

OmniaKey lists Hydrogen at its model page; Carbon runs through Google's paid Gemini API, which has no free tier.

## Sources checked

- [Google: Gemini 4 Hydrogen model page](https://ai.google.dev/gemini-api/docs/models/gemini-4-hydrogen) (checked November 13, 2026 UTC)
- [Google: Gemini 4 Carbon model page](https://ai.google.dev/gemini-api/docs/models/gemini-4-carbon) (checked November 13, 2026 UTC)
- [Google: What's new in Gemini 4 Hydrogen](https://ai.google.dev/gemini-api/docs/whats-new-gemini-4-hydrogen) (checked November 13, 2026 UTC)
- [Google: What's new in Gemini 4 Carbon](https://ai.google.dev/gemini-api/docs/whats-new-gemini-4-carbon) (checked November 13, 2026 UTC)
- [Google: Gemini thinking](https://ai.google.dev/gemini-api/docs/thinking) (checked November 13, 2026 UTC)
- [Google: Gemini API pricing](https://ai.google.dev/gemini-api/docs/pricing) (checked November 13, 2026 UTC)
- [Google: Gemini 4 Carbon is now available](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/) (November 10, 2026)
- [Artificial Analysis: Gemini 4 Hydrogen (high)](https://artificialanalysis.ai/models/gemini-4-hydrogen-high) (snapshot November 13, 2026 UTC)
- [Artificial Analysis: Gemini 4 Carbon](https://artificialanalysis.ai/models/gemini-4-carbon) (snapshot November 13, 2026 UTC)
- [Artificial Analysis: Terminal-Bench 4.0 method](https://artificialanalysis.ai/evaluations/terminalbench-4-0)
- [Artificial Analysis: FrontierCode-AA 2.0 method](https://artificialanalysis.ai/evaluations/frontiercode-2-0)

## Update log

- **November 13, 2026**: First version. Model IDs, thinking levels, limits and prices checked against Google's pages; scores from Artificial Analysis's November 13, 2026 snapshot. Recheck Carbon's price after the December 1 change.

Hydrogen is the model you can start on now with an OmniaKey key. See the [Gemini 4 Hydrogen model page](https://omniakey.com/gemini-4-hydrogen) for current capabilities and pricing.
