From coding to search, Gemini 4 Hydrogen prioritizes fast execution

From development to Search, Gemini 4 Hydrogen targets faster execution, with four thinking levels and dynamic routing between models.

A single coding task can now be automatically distributed across multiple models in Google Antigravity. The rollout of Gemini 4 Hydrogen accompanies the introduction of Orchestrator, which switches between planning and execution within a session. The new model is also available independently for coding, document analysis, and tasks involving tools. Its rollout extends to the Gemini app and Google Search.

The workflow described in the Antigravity announcement can begin with read-only exploration assigned to Carbon: examining dependencies, resolving ambiguities, and preparing a plan. Edits, updates, and tests are then handled by Hydrogen. If execution encounters a roadblock, the system can return to Carbon to revise the plan. Model selection therefore evolves as the work progresses, rather than being determined solely by the initial request.

Access depends on the subscription. Within Antigravity, Hydrogen is rolling out to Google AI Pro and Ultra subscribers. The Argon-based Orchestrator mode is available on both plans, while the Carbon-based version is reserved for Ultra. Enterprises can access it through Gemini for business. The provider also describes configurations in which Carbon coordinates multiple Hydrogen agents assigned to develop and verify changes in separate workspaces.

The model’s speed-focused positioning comes with four thinking levels: low, medium, high, which is the default, and xhigh. Artificial Analysis measurements at high give it a score of 51 on the firm’s Intelligence Index. The xhigh setting reaches 54. At launch, the lab compared these results with Argon’s score of 50. The index combines multiple evaluations of reasoning, coding, and tool use; the difference does not establish superiority on every task.

The additional effort has a measurable cost. Across the index’s evaluations, Artificial Analysis calculates a weighted average cost of $1.87 per task at high, compared with $3.16 at xhigh. Total output token use rises from 110 million to 304 million across the full evaluation suite. These volumes describe the testing campaign, rather than an individual response. Identical per-token pricing can therefore produce different bills depending on how much reasoning and text the model generates.

Compared with Gemini 3.8 Flash, the provider reports tasks completed 2.5 times faster at high and a Terminal-Bench 4.1 score of 57%, versus 18.2%. The described improvements include targeted repository edits, parallel file reads and searches, and resuming work when a tool fails. That speed factor applies to Google’s reported comparison, without guaranteeing the same improvement on every project.

Multimodal work covers PDFs, spreadsheets, charts, images, audio recordings, and videos. Published demonstrations range from creating a presentation based on a recorded keynote to building interactive web interfaces. Another demonstration assembles public-domain archival footage into a narrated edit using a video production tool. The model’s own output is text: these examples involve software tools and do not represent a new native video generator.

The API documentation clarifies an important limit behind the million-token announcement. Input, reasoning, and output share a single 1,048,576-token window. It is not one million input tokens plus another million output tokens. Synchronous responses are capped at 65,536 tokens per step; longer generations require background execution or continuations, within the remaining context capacity.

Standard pricing is $1.50 per million input tokens and $7.50 per million output tokens, including reasoning, with no announced surcharge based on context length. A free tier, subject to rate limits, is available in AI Studio and the Gemini API. Integration also extends to Android Studio, while availability in Sheets begins with eligible consumer and business subscriptions.

Reducing fabricated responses is another launch claim. The provider reports a reduction of more than fivefold compared with 3.8 Flash in its evaluation. The model card still lists hallucinations among known limitations and notes slowdowns, request timeouts, and increased token consumption at higher thinking levels. Its automated evaluations also report more unjustified refusals, particularly on multimodal prompts. A detailed safety and security report is expected in the following days.

Within the Gemini app, free access does not mean users can consistently select the model from the dropdown: a launch footnote specifies that free accounts receive access through automatic model selection when appropriate for the request. The announced improvements concern adherence to requested formats, extended conversations, and the use of connected apps to gather information needed for a task.

The Google Search rollout is announced globally for all users. Its query fan-out technique breaks a question into simultaneous searches across relevant subtopics, broadening the links examined when preparing an answer. Visible changes are expected to include more images, comparison tables, and tools tailored to the request, alongside improved language understanding across supported AI Mode languages.