A model upgrade can change a coding agent’s bill even when you send it the same repository and ask it to fix the same bug.
Google released Gemini 4 Hydrogen on November 12 for coding, document work, and agents that call tools. Developers can use it through Google AI Studio or the Gemini API, where Google lists a rate-limited free tier and standard paid rates of $1.50 per million input tokens and $7.50 per million output tokens. The output rate includes billable reasoning. Google recommends Hydrogen as an upgrade from Gemini 3.8 Flash, which still appears on the developer price list.
Updated .
How much Gemini 4 Hydrogen costs
The Gemini Developer API price list shows Gemini 3.8 Flash at $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. Both rates double on January 1, 2027, matching Hydrogen’s standard rates. The example below holds usage steady at 20,000 input tokens and 2,000 billed output tokens per task.
| Model and rate period | Input / output per 1M | Cost per task |
|---|---|---|
| 3.8 Flash through Dec. 31 | $0.75 / $3.75 | $0.0225 |
| 3.8 Flash from Jan. 1 | $1.50 / $7.50 | $0.0450 |
| 4 Hydrogen | $1.50 / $7.50 | $0.0450 |
At the listed rates, the same 10,000 tasks would cost $225 on Flash during the promotion and $450 on Hydrogen. Hydrogen would need to cut the weighted token cost of the Flash workload in half to match the promotional bill. Input and output tokens have different prices, so dropping 50% of input tokens alone wouldn’t be enough. For this example, halving only the input brings Hydrogen to $0.0300 per task. Halving both input and billed output brings it to $0.0225.
Another way for the higher rate to pay off is through higher throughput. If Hydrogen completes more tasks without retries or a second model, it can reduce the cost of getting a usable result. The figures above are calculated from the rate card. They exclude caching, grounding, and retries, and they do not measure either model’s success rate.
Thinking levels, output limits, and caching
Google’s developer guide lists four thinking levels: low, medium, high, and xhigh. The default is high. For a latency-sensitive tool call, start by comparing medium with the default. For an issue that requires changes across several files, try xhigh on cases the default struggled to complete. Hold the other settings steady so you can see which change affected the result.
The 1,048,576-token window is shared by input, reasoning, and output. Google documents a 65,536-token synchronous output limit per hop and supports longer generations through background execution or client-side continuation. Your own max_output_tokens setting also covers reasoning and the visible response. If an agent stops halfway through a patch, log how much of its output allowance it used before assuming it lost track of the files.
The migration guide lists two other changes worth testing. Hydrogen uses implicit context caching, with eligible cached input billed at $0.075 per million tokens on standard paid requests. If your Flash integration creates an explicit cache and passes its ID, remove that field in the Hydrogen path and measure cache hits. Google also says Hydrogen ignores custom temperature, top_p, and top_k values. An evaluation that relies on those settings should rerun its baseline.
A migration test for coding teams
Choose a small set of issues with known acceptance tests. Include a one-file fix, a change that spans several modules, and a task with a tool failure the agent must recover from. Run both models against identical repository snapshots and keep tool access and acceptance criteria fixed. Record billed input, cached input, output, retries, fallback calls, and total time for each attempt.
Divide total spend by accepted patches and report the acceptance rate beside it. Check the slowest runs as well as the median. That exposes regressions a short demo can miss, such as a tool loop that passes the tests but keeps users waiting. Google’s model card documents its own evaluations. The workflow here is a proposed test for your repository. If the task only classifies an input, compare the workflow with the separate Gemini Decisions API, which prices a shared input differently.
Where else Google is rolling out Hydrogen
Google says Hydrogen also powers an upgrade to Search for complex questions, changing how it retrieves links and presents results. Free Gemini app users can get Hydrogen through automatic model selection when Google chooses it for a prompt. Google lists Antigravity access for paid Google AI Pro and Ultra subscribers. Teams looking to call a specific model in production can select gemini-4-hydrogen through the API.
Abdessalam Alaoui
I'm Abdessalam Alaoui, founder of Musthave.AI. I write practical guides to AI models, tools, generative media.
