Gemini 4 Hydrogen vs Gemini 3.8 Flash, Luna, Terra, Haiku and Sonnet

Google positions Gemini 4 Hydrogen as the upgrade from Gemini 3.8 Flash, but at November 2026 list prices its standard API rate is twice Flash's introductory rate: $1.50 per million input tokens and $7.50 per million output tokens. Flash keeps that introductory rate through December 31, 2026.

Hydrogen costs more per token

Google announced Hydrogen on November 12, 2026. Hydrogen's standard price applies across its context window, while Flash's introductory pricing runs through the end of the year.

Sources: Google's Hydrogen announcement, Google Cloud pricing and Requesty's coding benchmark. Prices are November 2026 list rates per million tokens, except for Flash's January 2027 row.

At those rates, Hydrogen's input and output tokens each cost 15 times as much as Haiku's and Luna's. The cost of a completed task also depends on how many tokens each model uses and whether it returns a usable answer.

Google Cloud's Batch tier halves Hydrogen's price to $0.75 per million input tokens and $3.75 per million output tokens. On Requesty, the Flex endpoint google/gemini-4-hydrogen:flex is listed at the same rates.

What changes with Hydrogen?

The Artificial Analysis comparison reports high-setting intelligence index scores of 51 for Hydrogen and 31 for Gemini 3.8 Flash.

Google reports Terminal-Bench 4.1 scores of 57% for Hydrogen and 18.2% for Flash, and 71.6% for Hydrogen on Agents' Last Exam v1.1. These are vendor figures, separate from Requesty's coding benchmark.

Google also reports that Hydrogen finishes tasks 2.5 times faster than Flash at the default (HIGH) thinking level. Like the benchmark scores above, this is a vendor claim.

Hydrogen has four thinking levels: LOW, MEDIUM, HIGH and XHIGH. Google's documentation lists HIGH as the default, a unified input plus output limit of 1,048,576 tokens, and up to 65,536 output tokens in one foreground segment.

Output billing includes thinking tokens. Include those tokens when comparing the cost of a task, even if they are separate from the answer shown to your user.

What we measure

Our coding benchmark gives a controlled comparison of the alternatives. Each model ran on the same 129 private Go backend tasks with the same agent and an 80 turn budget.

Sonnet had the highest solve rate of these four models, while Luna had the lowest cost per solve. Haiku solved more tasks than Luna at twice the cost per solve. The benchmark notes that some providers do not count hidden reasoning in output tokens, so GPT throughput can read lower than the work done.

Requesty Rankings adds a view from production traffic between September 7 and November 9, 2026. Time to first token (TTFT) is the wait before a streamed response starts; the speed figures below also come from streamed responses. Hydrogen launched after this window, so it has no production figures yet.

The blended cost is what we billed after cache discounts. Workload and cache mix affect that number. For more on cached input prices, see our prompt caching comparison.

The break-even point for a Flash migration

Take a set of tasks that uses 1 million input tokens and 200,000 output tokens on Gemini 3.8 Flash. At Flash's introductory standard rates, input costs $0.75 and output costs $0.75, for a total of $1.50.

Say each alternative completes the same set of tasks with the same input and output token counts. Using the November 2026 list rates above, without cache discounts, the bills look like this:

For Hydrogen to match Flash's $1.50 bill at standard rates, its input and output counts must satisfy this condition:

1.50 × input tokens in millions + 7.50 × output tokens in millions = $1.50

One example is 500,000 input tokens and 100,000 output tokens. Other mixes can reach the same bill. Retries or fewer usable answers change the cost per usable answer.

From January 1, 2027, Flash's listed rates match Hydrogen's. The same token counts would then have the same standard token cost.

The coding benchmark shows why token price is only part of the bill: Luna's $0.07 per solve came with a 65.1% solve rate, while Haiku's $0.14 came with an 81.4% solve rate. Our coding model comparison has more benchmark results.

Cached input changes the bill

Hydrogen supports implicit caching for prompts of at least 4,096 tokens, according to Google Cloud's model documentation. Cached input costs $0.075 per million tokens, 95% less than the standard $1.50 input rate.

For a workload that repeats a long prompt, the cached portion of the input can cost less than the standard price suggests. Keep cached and uncached input separate in your calculation, and include the output tokens.

What to test before switching

Run a sample of your Flash tasks through Hydrogen and record input tokens, output tokens and usable answers for each model. Test the thinking level you plan to use in production: HIGH is Hydrogen's default, and thinking tokens count toward output billing.

For more context on the OpenAI models, see our guide to GPT-6 Sol and Luna pricing.

If you want to test models through one API, Requesty's AI gateway has an OpenAI-compatible endpoint for models from multiple providers, including Hydrogen as google/gemini-4-hydrogen. Start with a fixed set of tasks, record the full token bill for each run, and use those results to decide which model earns a place in production.