Google shipped Gemini 4 Hydrogen on November 12, 2026, and on the Artificial Analysis index it lands one point behind Claude Sonnet 5.5 (54 vs. 55, both at their highest reasoning setting) at roughly $3.16 per task, about a third of what Sonnet costs. It replaces the Flash line, it's on the free tier of the API, and at $1.50 per million input tokens it's 25% cheaper per token than Sonnet 5.5.
What Hydrogen brings
Google pitches it as the fast executor of the Gemini 4 family: Carbon and Argon handle long-horizon reasoning, while Hydrogen writes code, calls tools and drives agents. The surprise is that at its default high level it already scores 51 on the Artificial Analysis index, ahead of Argon (50), which launched less than four weeks earlier.
| Spec | Gemini 4 Hydrogen |
|---|---|
| Model ID | gemini-4-hydrogen |
| Paid-tier price | $1.50 / $7.50 per million input / output tokens |
| Cached input (implicit) | $0.075 per million (95% off) |
| Context | 1,048,576 tokens shared by input, thinking and output |
| Output per request | 65,536 tokens; up to 1M with background execution |
| Thinking levels | low, medium, high (default) and the new xhigh |
Against Gemini 3.8 Flash, Google claims tasks finish 2.5x faster at high, Terminal-Bench 4.1 jumps from 18.2% to 57%, and fabricated claims drop fivefold. Artificial Analysis backs up the last point: on AA-Omniscience, Hydrogen makes things up only 8% of the time.
xhigh isn't free. Artificial Analysis found it burns almost twice the output tokens of high, pushing cost per task from $1.87 to $3.16 for a three-point gain (51 to 54). Below that, medium costs $0.80 per task and low $0.29, but the index drops to 41 and 27.
Trying it
It's available in the Gemini API (both generateContent and the Interactions API) and in Google AI Studio. With the Python SDK and an explicit thinking level:
from google import genai
client = genai.Client()
interaction = client.interactions.create(
model="gemini-4-hydrogen",
input="Audit this payment retry queue for race conditions.",
generation_config={"thinking_level": "high"},
)
print(interaction.output_text)Coming from Gemini 3.8 Flash, swapping the model ID isn't enough. Google's migration guide lists changes that break existing requests:
thinking_level: "minimal"now returns an error, and the default moves frommediumtohigh, so your bill goes up unless you pin it.- Explicit caching is gone: drop
cached_contentandcaches.create; implicit caching kicks in automatically for prompts of 4,096 tokens or more. temperature,top_pandtop_kare ignored.- Tool calls must be declared through function calling or structured outputs; asking for hand-rolled JSON in the prompt can end in
MALFORMED_FUNCTION_CALL. - Output per request caps at 65,536 tokens; any low
max_output_tokensyou hardcoded will truncate long answers.
Against Claude Sonnet 5.5
Sonnet 5.5 is the benchmark in this tier: Anthropic's mid-size model, with the same 1M context and a very similar score.
| Gemini 4 Hydrogen | Claude Sonnet 5.5 | |
|---|---|---|
| Input / output price (1M tokens) | $1.50 / $7.50 | $2 / $10 |
| Cache read (1M tokens) | $0.075 | $0.10 |
| Context | 1M (shared) | 1M |
| Max output per request | 64K (1M in background) | 128K (300K on Batch, beta) |
| Artificial Analysis index (max) | 54 (xhigh) | 55 |
| Terminal-Bench 4.0 (AA) | 54% | 64% |
| Terminal-Bench-Science 0.1 (AA) | 50% | 53% |
| Cost per task (AA, max level) | $3.16 | about three times as much |
The per-token gap is only 25%; the rest of the cost-per-task difference comes from how many tokens each model spends on the same job.
Where it wins and where it loses
Hydrogen wins on price and on knowledge work: long documents, PDFs, spreadsheets and video, where Google's own charts put it ahead of Sonnet 5.5 and where it rarely hallucinates. It also fits as the workhorse in multi-agent setups: Google positions it as the worker and verifier behind Carbon, and the Antigravity orchestrator that switches between the two gets close to Carbon's Terminal-Bench 4.0 success rate on 24% less quota.
It loses where it hurts most if you automate in a shell. Ten points behind Sonnet 5.5 on Terminal-Bench 4.0 is a lot for an agent that runs commands, reads errors and retries, and Google's own chart concedes that Sonnet still leads on terminal and computer use. It isn't the cheapest option either: Artificial Analysis puts it at about 2.5 times the cost per task of GPT-6.1 Sol, which scores 56 and 58 on the two terminal benchmarks.
Verdict
If you're on Gemini 3.8 Flash today, Hydrogen is the obvious upgrade, with the migration guide open next to you. If your workload is documents, extraction, or agents that look things up and summarize, it deserves a serious trial against Sonnet 5.5: nearly the same quality, and the bill could shrink to a third. If your agent lives in a terminal, Sonnet 5.5 is still the safer bet, and before switching, measure cost per task on your own workload rather than price per token.
Sources
- Introducing Gemini 4 Hydrogen, Google, November 12, 2026.
- What's new in Gemini 4 Hydrogen, Gemini API docs.
- Google's Gemini 4 Hydrogen outscores the bigger Argon just four weeks after its launch, The Decoder, November 12, 2026.
- Claude Sonnet 5.5, Anthropic docs.