Google launches Gemini 4 Hydrogen — a faster, smarter, pricier Flash successor

Google released Gemini 4 Hydrogen on Thursday, giving developers a faster model for writing code, working through documents and running AI agents that take several steps to finish a job.

Google says Hydrogen completes tasks 2.5 times as fast as Gemini 3.8 Flash at Hydrogen’s default reasoning setting. In Artificial Analysis’s independent testing, Hydrogen at that setting scored 51 on the firm’s Intelligence Index, up from 31 for Flash at high reasoning and a point ahead of Gemini 4 Argon at high.

That puts Hydrogen in an unusual spot. It can take on some of the work teams might send to a more expensive reasoning model, but it costs 15 times as much per uncached input and output token as the lowest standard rates for Anthropic’s Claude Haiku 5.5, OpenAI’s GPT-6 Luna and Meta’s Muse Ember.

Google lists standard paid Gemini API rates of $1.50 per million uncached input tokens, $0.075 per million cached input tokens, and $7.50 per million output tokens, including reasoning tokens. Google lists no introductory promotion for Hydrogen’s standard tier.

Batch, Flex, and Contributor each cost half the standard rates. Batch processes groups of requests in the background, with a target turnaround of 24 hours. Flex handles individual requests at lower priority, with a target response time of 1–15 minutes and limited availability when demand is high. The invite-only Contributor tier offers standard-tier speed and reliability in exchange for letting Google use data from participating requests to improve its products and train models.

For a company deciding which model to put behind a coding assistant, document processor or customer support tool, the question is how often Hydrogen’s extra capability saves a retry, catches an error or gets a job done that a cheaper model would miss.

What changes for Gemini 3.8 Flash users

One of the biggest jumps Google reports is in work that requires an agent to operate a terminal. On Google’s Terminal-Bench 4.1 evaluation, Hydrogen completes 57% of tasks, up from 18.2% for Flash in Google’s launch comparison. It still trails Claude Sonnet 5.5’s 63.9% in Google’s table.

Google also reports that Hydrogen recovers more reliably when a tool call fails or a search comes back empty. In a codebase, that can mean trying another search or reading a different file after the first attempt fails.

In an early-access test, Box’s agent went from 74% accuracy with Flash to 79% with Hydrogen on its Complex Work Eval, according to a chart Box CEO Aaron Levie posted on X. Levie said Hydrogen took less than half as long and used about 12% fewer tokens on a typical task.

Box’s AI product team told Google that it saw the largest improvements on work that required reading tables and doing calculations. Levie said Box AI Studio support was coming. The speed and accuracy gains tell one part of the story. The bill also depends on how many of those tokens are cached and how much work the model has to repeat.

For developers moving an existing application, Google’s migration guide lists changes beyond the model name. Hydrogen defaults to high reasoning, up from medium on Flash, and adds an xhigh option. It rejects the old thinking_budget parameter and the MINIMAL setting.

On Google Cloud, automatic caching starts at 4,096 tokens, down from 6,144 on Flash, and explicit caching is unavailable at launch. Hydrogen accepts text, images, audio and video, and Google Cloud can ingest Office and OpenDocument files directly. Its roughly 1 million-token window is shared by the input and output, including reasoning tokens and tool calls. Supported APIs can continue generating past the 65,536-token limit of a single foreground segment.

How Hydrogen compares with Haiku, Luna and Ember

Google charges $1.50 per million input tokens and $7.50 per million output tokens at its standard global rate, with cached input at $0.075 per million. The rates are flat across the context window. That matches Flash’s scheduled standard pricing from January 1, 2027, and is twice Flash’s introductory rate through December 31. Google’s deprecation schedule lists no shutdown date for Flash.

The table uses the Artificial Analysis Intelligence Index v5.0.1, which combines nine evaluations across areas including coding, document analysis and science. Cost per task is the firm’s weighted estimate for that suite. API rates are for uncached input and output and can vary with prompt length, provider or time of day.

Even at medium, Hydrogen outscored Haiku, Luna and Ember at max, at a higher estimated cost per task than any of the three. The advantage shows up in some coding tests. In Artificial Analysis’s separate Terminal-Bench 4.0 runs, Hydrogen scored 54% at both high and xhigh, compared with 40.9% for Ember and 12.6% for Luna and 32.8% for Haiku, all three at max.

On another coding test, FrontierCode 2.0, Haiku beat Ember 34.9% to 27.2%. For teams already paying for Hydrogen, the higher reasoning setting bought no improvement on Terminal-Bench 4.0 while lifting the estimated cost across the full suite from $1.87 to $3.16.

The three U.S. rivals also differ in what they accept and where they run. Haiku takes text and images and is available via Anthropic, Amazon Bedrock, Google Cloud and Microsoft Foundry. Its API rates rise to $0.50 input and $2.50 output for prompts over 100,000 tokens.

Luna takes text and images through OpenAI’s API, with rates of $0.20 input and $0.75 output for prompts over 272,000 input tokens. Meta’s own Ember API takes audio and video as well as text, images and PDFs.

Ember is also in Muse Code, and Google Cloud offers a text-and-image preview. Meta’s rates rise to $0.30 input and $1.50 output above 200,000 input tokens.

Where China’s open-weight models fit

DeepSeek V4.1 Flash 1017 offers text and image input, a million-token window and MIT-licensed weights. DeepSeek’s hosted API drops to $0.15 input and $0.60 output off-peak. Xiaomi’s MiMo-V2.6-Flash also offers MIT-licensed weights, with audio and video support documented for the downloadable model.

Artificial Analysis puts Xiaomi’s hosted Flash model at $0.13 per benchmark task, versus $1.87 for Hydrogen at high, with a 24-point gap in the composite score. The open weights give companies a choice of running the models themselves or buying hosted access, with the associated hardware and operating costs for self-hosting.

Moonshot’s Kimi K3.1 Pro is closer on capability. It scores 46 at an estimated $3.68 per task through Kimi’s API. Its weights come with Moonshot’s own license, and multiple providers host it. That gives buyers more deployment options even where Google’s hosted model has the edge on this particular benchmark.

Where to get Hydrogen

Hydrogen is rolling out globally via the Gemini API, AI Studio, Gemini for business, the Gemini app and Google Search. Free Gemini app users get it through automatic model selection. Individual Antigravity access requires Google AI Pro or Ultra. Google Cloud lists gemini-4-hydrogen as generally available, and Google says eligible business subscriptions also get it in Gemini in Sheets.

Google is also building the choice between models into Antigravity. Its new Orchestrator can switch between Carbon and Hydrogen in the same session.

Google reports that the combined setup came close to Carbon’s success rate on Terminal-Bench 4.0 while finishing 19% faster and using 24% less quota.

That puts Hydrogen in play for two different jobs: taking over work from Flash and handling some of the work a larger model would otherwise do. The savings depend on which model the work would have gone to in the first place.