Gemini 4 Hydrogen brings four thinking levels to everyday coding

Google released Gemini 4 Hydrogen on November 12 for tasks such as editing code, running tests and checking the results. Developers can use it directly through the Gemini API or let Google Antigravity’s new Orchestrator switch models during a session. The immediate decision for a team is how much reasoning time to spend on each step of a job. Google’s launch announcement

How much does a Hydrogen task cost?

Google lists $1.50 per million input tokens and $7.50 per million output tokens, with the same rates across the million-token context window. At those rates, a request with 100,000 billed input tokens and 10,000 billed output tokens costs $0.225 before other charges. The shorter output accounts for one-third of that example bill.

That arithmetic matters for repository work. If a patch fails and an agent reads the same files again, the team pays for another attempt. For a migration trial, record the total billed tokens through the first passing test. That measures the cost of the completed change.

Which thinking level should a team try first?

Hydrogen defaults to high and also supports low, medium and xhigh. Start with the default on a real maintenance ticket whose acceptance test already exists. Then repeat the ticket at a lower setting with the same files and tools. Compare the patch, test result, elapsed time and cost. A fast response that requires a second developer to finish the edit changes the economics.

For a product such as AI Web Awards, one possible evaluation would be updating a scoring report against a fixed set of example submissions. A reviewer could compare the generated report against expected scores before approving a change. This is a proposed test, not a report of Hydrogen running in that product.

Where do Carbon and Orchestrator fit?

Google positions Gemini 4 Carbon for harder planning and Hydrogen for faster execution. In Antigravity, Orchestrator can switch between them within one session. Google reports 24% lower average quota use and 19% faster completion on its Terminal-Bench 4.0 comparison with Carbon. Those figures describe that benchmark. API token prices and Antigravity subscription quotas measure different things.

How can developers access Hydrogen?

Google lists AI Studio and the Gemini API across tiers. It also lists the Gemini app and Search, while Antigravity access at launch is for paid Google AI Pro and Ultra subscribers. Teams already using Vercel can evaluate the model through AI Gateway.

If you are budgeting a rollout, begin with a week of representative tasks and keep the existing model available during the trial. FindMilan’s AI consulting service can help define the acceptance tests and compare cost per completed job across providers.