Overview
Gemini 4 Hydrogen is a good first model to test for routine repository edits, test writing, and coding assistants that call tools. Google released it on November 12, 2026, as gemini-4-hydrogen on the free and paid Gemini API tiers. If you already use Gemini 3.8 Flash, run Hydrogen on representative jobs before switching production traffic. The models budget context differently, and Hydrogen can cost more at Flash's introductory rates.
What Changed
- Four thinking levels. Hydrogen adds
xhightolow,medium, andhigh. The API now defaults tohighfor this model. Pin a level when comparing latency and cost. - One shared context window. Input, thinking, and output share 1,048,576 tokens. Ordinary single requests cap thinking and output at 65,536 tokens. Longer responses require background execution or client continuation and still share the context budget.
- A new Managed Agents default. Google switched the Antigravity agent in Gemini Managed Agents to Hydrogen. Review the configured model if you depend on a specific model's behavior.
- Free and paid developer access. Hydrogen is available through GenerateContent and the Interactions API. You can also try it in Google AI Studio.
Compared With Previous Version
On 3.8 Flash, input and output have separate limits. On Hydrogen, a long input leaves less room for reasoning and output. If your app feeds a whole repository or a large document bundle to the model, reserve headroom before sending the request.
Flash defaults to medium thinking. Test Hydrogen at medium and high on the same set of jobs, then compare the share of jobs that pass your checks, median latency, and cost per successful job. Higher thinking effort can increase cost even when the visible answer stays short.
Migration checks
- Try
gemini-4-hydrogenon a sample of requests and set the thinking level explicitly. Hydrogen rejectsminimal. - Replace explicit cache references. Hydrogen supports implicit caching, so verify actual cache hits and costs on your repeated prompts before switching.
- Remove sampling controls you rely on, such as
temperature,top_p, andtop_k. The API ignores custom values on Hydrogen. Retest output consistency and any structured responses your application parses. - Recheck limits on long prompts and test your tool calls over multiple turns. Use Google's migration checklist for API-specific details.
Pricing and Access
These are Google's Standard paid Gemini API rates in USD per million tokens as checked on November 13, 2026. Output rates include thinking tokens.
| Model | Input / output | Free API tier |
|---|---|---|
| Gemini 4 Hydrogen | $1.50 / $7.50 | Available |
| Gemini 3.8 Flash | $0.75 / $3.75 through Dec 31, 2026, then $1.50 / $7.50 from Jan 1, 2027 | Available |
| Gemini 4 Argon and Carbon | $2 / $10 through Nov 30, 2026, then $4 / $20 from Dec 1, 2026 | Paid access |
At the listed Standard rates, Hydrogen costs twice as much per uncached input or output token as 3.8 Flash through December 31. Their scheduled rates then match, but actual spend depends on how many thinking tokens each job uses, whether the input hits the cache, and any separate tool charges. Gemini API pricing
For everyday chat, Google says free Gemini app users can receive Hydrogen through automatic model selection. In the Antigravity development app, Google lists access for AI Pro and Ultra subscribers. Those access rules are separate from Gemini API billing. Google launch announcement
Who Should Upgrade and Who Should Wait
Try Hydrogen for everyday code changes, test fixes, and tool-driven document tasks where quick feedback matters. Start with a small evaluation that checks completed work and records the cost of retries.
Keep 3.8 Flash in production while you test if your app depends on explicit caching, fixed sampling behavior, or Flash's introductory rates. A model ID swap alone can change both behavior and spend.
Try Argon or Carbon when difficult tasks consistently need more reasoning. Argon is a candidate for multi-step engineering work. Reserve Carbon for hard architecture decisions or debugging cases where extra verification is worth the longer run. For a mixed workload, test the stronger model on the hardest decisions and let Hydrogen handle routine edits and tests.
