Skip to content
All insights

AI & Automation

Gemini 4 Carbon vs Hydrogen: Benchmarks and API Cost

Compare Gemini 4 Carbon and Hydrogen on benchmarks, effort levels, and December API pricing. Work through a budget for routing difficult jobs to Carbon.

Start with Gemini 4 Hydrogen for frequent, well-defined tasks that have cheap automated checks. Test Gemini 4 Carbon on the difficult cases that fail those checks. Google opened Carbon to broader public access on November 10, 2026, and introduced Hydrogen on November 12. Hydrogen costs less per token on Google’s standard paid Gemini Developer API, and Carbon’s introductory rates end after November 30.

Here’s how their benchmark scores change with reasoning effort and what the December price change means for a product that routes exceptions to Carbon.

Key takeaways

  • Google lists gemini-4-carbon as paid-only and documents a free tier for gemini-4-hydrogen.
  • Artificial Analysis scored Carbon High at 60 and Hydrogen Xhigh at 54 on its Intelligence Index v5.0.1. Hydrogen Xhigh and Carbon Low both scored 54.
  • From December 1, Carbon’s standard paid input and output rates rise to $4 and $20 per million tokens. Hydrogen’s published rates are $1.50 and $7.50.
Applying this to your product?Plan your model routing test

Where to try Carbon and Hydrogen

Google’s Carbon announcement lists developer access through the Gemini API and consumer access in the Gemini app and AI Mode for Google AI Ultra subscribers. Hydrogen is available in the Gemini API and across Google products including the Gemini app and Search. App entitlements and API billing have separate controls. Confirm the model ID and tier from the project where your integration will run.

Google’s Carbon public release

Google’s Hydrogen release and availability

Gemini 4 Carbon vs Hydrogen benchmarks

Carbon High leads the four configurations below in Artificial Analysis’s composite index. Hydrogen Xhigh and Carbon Low both display 54, but Hydrogen Xhigh scored higher on AutomationBench-AA and lower on Terminal-Bench 4.0. AutomationBench-AA measures tool-based workflows. Terminal-Bench tests tasks in a terminal. Use the individual results to choose which effort levels to pilot.

For catalog intake, accuracy means extracting the right units, keeping each price attached to the right SKU and flagging gaps. Use direct checks for those outcomes, along with a deadline that matches the product’s actual latency budget.

Artificial Analysis results by thinking level, checked November 18, 2026
Model and effortIntelligence Index v5.0.1AutomationBench-AATerminal-Bench 4.0
Carbon High6077.8%65.2%
Carbon Low5463.5%62.6%
Hydrogen Xhigh5476.1%53.5%
Hydrogen Medium4172.0%29.3%

Artificial Analysis comparison by thinking level

Gemini 4 Carbon and Hydrogen API pricing

These prices apply to standard requests on Google’s paid Gemini Developer API. Output billing includes thinking tokens. Google also lists implicit cache rates and separate charges for some tools. A budget for long documents should use recorded billable tokens and count the full cost of reruns.

Standard Gemini Developer API rates in USD per million tokens
Model and periodInputOutput including thinking
Carbon through November 30, 2026$2.00$10.00
Carbon from December 1, 2026$4.00$20.00
Hydrogen, checked November 18, 2026$1.50$7.50

Google’s official Gemini Developer API pricing

FROM READING TO DOING

Which jobs should reach the costlier model?

Bring one representative workflow, its failure cases and the cost or latency your team can accept. We can help test a routing rule on real inputs.

Worked example for routing ambiguous catalog files

Suppose a commerce SaaS imports 1,000 supplier files. Each attempt uses 10,000 uncached input tokens and 4,000 billed output tokens, including thinking. Hydrogen handles every file first and 20% of the files require a full Carbon rerun. That makes 1,000 Hydrogen attempts plus 200 Carbon attempts. The share escalated is an assumption. The model can use different token counts in a real rollout.

Illustrative model charges for 1,000 files in USD, before tools and review
StrategyNovember ratesDecember rates
Carbon on every file$60.00$120.00
Hydrogen first, then 200 Carbon reruns$57.00$69.00

Model cost per accepted AI outcome

Test routing errors before you automate decisions

Build an evaluation with known SKU mappings, duplicate rows, missing prices and conflicting currencies. Validate fields independently of the model. Escalate contradictory rows or missing evidence when the first pass hits a documented rule. Keep unresolved outcomes in a review queue so an uncertain result can’t silently update a catalog.

Record how many erroneous results passed the first gate, how many good results escalated, time to an accepted import and true cost per accepted file. Compare those outcomes with your current setup on the same fixtures. Sample accepted Hydrogen results during rollout to check for failure modes your rules missed.

Check the connector and job lifecycle in the Gemini 4 API migration guide

Frequently asked questions

Is Gemini 4 Hydrogen free?

Google lists a free tier for Hydrogen on the Gemini Developer API. Check that tier’s availability, limits and data terms before using it for a product. The pricing example uses standard paid API rates.

When does Gemini 4 Carbon’s API price go up?

Google lists December 1, 2026, for the standard paid Carbon rate change from $2 input and $10 output to $4 input and $20 output per million tokens. The output rate includes thinking tokens.

YOUR NEXT STEP

Let’s work through your next step.

Tell us what you’re building, what you’ve tried, and where you need a hand. We’ll work with you to define a practical way forward.

Keep reading.

Choose which optional tools can run. You can reopen these settings from the footer. These choices do not subscribe you to marketing emails.

Cookie Policy