Google Releases Gemini 4 Hydrogen as Fast and Cheap Flagship

宝玉 (@dotey) · 2026-11-12 18

Google releases Gemini 4 Hydrogen: the flagship model benchmarked against Claude Sonnet 5.5 and GPT-6.1 Terra, available to Gemini App paid users, and usable on the developer free tier

Google released Gemini 4 Hydrogen on November 12. It is the fast-and-cheap tier of the Gemini 4 series, succeeding Gemini 3.8 Flash, and Google positioned it against Anthropic's Claude Sonnet 5.5 and OpenAI's GPT-6.1 Terra at launch.

The division of labor among the three Gemini 4 models: Carbon is the most capable, reserved for the hardest and longest tasks; Argon handles most complex tasks and uses fewer tokens than Carbon; Hydrogen handles everyday coding and agent runs, and can also assist Carbon.

In the Artificial Analysis composite index published by Google, Hydrogen scores 51 in its default tier, slightly above Argon's 50, and 54 at its highest tier; Carbon scores 60 on the Artificial Analysis website.

A day earlier, Business Insider reported that Google employees were testing a model codenamed Hydrogen on the internal development platform Jetski, with one engineer saying in an internal message that it was "very strong, fast and cheap." As with Carbon, the internal codename became the official name.

[Compared with Sonnet 5.5 and GPT-6.1 Terra]

API list price, per million tokens input / output: Hydrogen $1.5 /$7.5, Claude Sonnet 5.5 $2 /$10, GPT-6.1 Terra $1 /$5. Hydrogen is three-quarters the price of Sonnet 5.5 and half again as expensive as Terra.

Of the 10 benchmarks Google published, Hydrogen ranks first on 7. The largest gap is on SWE-sweep, which requires finding and fixing bugs without breaking existing code: Hydrogen 13.5%, Sonnet 5.5 3.1%, Terra 5.8%. Sonnet 5.5 won the other 3: command-line coding on Terminal-Bench 4.1 (63.9% vs. 57.0%), computer operation on OSWorld (83.9% vs. 77.6%), and scientific tasks.

Factoring in the cost per task makes the picture more nuanced. According to Google's own chart, with all three at their highest tier, Hydrogen scores 54 at about $3 per task; Sonnet 5.5 also scores around 54 at about $10 per task; Terra scores 51 at under $1 per task. At the same score of 51, Terra is half the price of Hydrogen's default tier, but Hydrogen takes less than half the time of Terra.

[Who can use it, and whether it costs money]

Gemini App: Plus ($4.99 per month), Pro, and Ultra users can manually select Hydrogen in the model menu. Free users cannot select it and can only leave "automatic model selection" on; when a question needs a bit more thinking, the system may hand it to Hydrogen, while most questions still use the lightest Flash-Lite. With automatic selection off, everything uses Flash-Lite.

Google Search now runs on Hydrogen, available to all users. In Google's AI coding tool Antigravity, paying Pro and Ultra users can use it.

Developers: Hydrogen is available on the free tier of AI Studio and the Gemini API, rate-limited by call frequency; neither Argon nor Carbon is on the free tier. The paid tier is the $1.5 /$7.5 above, with no introductory discount, and batch and Flex tiers get another 50% off.

Compared with its own models, Argon and Carbon's introductory price of $2 /$10 runs through November 30 and rises to $4 /$20 from December 1, at which point Hydrogen's unit price will be less than 40% of theirs. 3.8 Flash's current promotional price of $0.75 /$3.75 runs through the end of the year; switching from 3.8 Flash today will first double the unit price, and from January next year the two will be level.

[What it can do]

Google says that compared with 3.8 Flash, Hydrogen completes tasks 2.5 times faster, raises Terminal-Bench 4.1 from 18.2% to 57%, cuts fabricated content by more than 80%, and can issue several operations such as reading files and searching in parallel at once.

Google recommends pairing Hydrogen with Carbon: Carbon does planning and tackles hard problems, while code changes and test runs go to Hydrogen. Antigravity has added an Orchestrator that automatically switches between the two models within the same session. Google's tests found success rates close to using Carbon throughout, with 24% less quota used and 19% faster.

In addition to low, medium, and high, the thinking tiers add xhigh, with high as the default; the context window is 1 million tokens, shared across input, thinking, and output.

The benchmark results and costs above are all published by Google itself; the Artificial Analysis website currently lists Hydrogen but does not yet have its score; the scores for Sonnet 5.5 (54.5 at its highest tier) and Terra (51.3) can be verified on the website.

Source: Google's official blog "Introducing Gemini 4 Hydrogen" (Koray Kavukcuoglu, 2026-11-12)

↳ Quoting @sundarpichai (Sundar Pichai): Following Carbon + Argon, Gemini 4 Hydrogen brings cutting-edge intelligence to everyone. Built for speed and efficiency for everyday tasks, coding and agentic workflows.

Compared to 3.8 Flash, it finishes tasks 2.5x faster. Also has 4 thinking levels, so you only wait and pay for the reasoning depth a task requires.

Rolling out today to all tiers of @GoogleAIStudio, the Gemini API, @GeminiApp, Search and paid Google AI Pro and Ultra subscribers in Google @Antigravity.