Google DeepMind has introduced Gemini 4 Hydrogen, a model aimed at everyday coding and agentic workflows, and says it completes tasks 2.5 times faster than Gemini 3.8 Flash at the default high thinking setting. The release also brings a new multi-model Orchestrator to Google Antigravity, which Google says can reduce quota use and completion time while approaching its larger model’s task success rate.
Google DeepMind Watch analysis
What happened
Hydrogen is available across Google AI Studio, the Gemini API, Gemini for business, the Gemini app and Google Search, according to Google DeepMind. In Antigravity, it is also available to paid Google AI Pro and Ultra subscribers. Google prices it at $1.50 per million input tokens and $7.50 per million output tokens across its one-million-token context window.
Google positions Hydrogen as a faster execution model that can work alone or alongside its larger Gemini 4 models, Carbon and Argon. Its new Antigravity Orchestrator can switch models during a session. Google says that on Terminal-Bench 4.0, Orchestrator achieved task success rates near Carbon’s while using 24% less quota on average and finishing 19% faster. Those are company-reported benchmark results, not independent comparisons.
Google also says Hydrogen more than tripled Gemini 3.8 Flash’s Terminal-Bench 4.1 score, from 18.2% to 57%, and cut fabricated claims by more than five times. For multimodal work, the company describes support for up to one million tokens of input and output, including synthesis of PDFs, spreadsheets, charts, audio and recorded video into cited reports. Google’s announcement gives the full model and benchmark details.
Why it matters
Hydrogen’s pitch is not simply “bigger model, bigger bill”. Google is aiming at the routine work around coding and agents, where speed, token use and the ability to recover from a failed tool call can matter more than having the most powerful model in every step. Orchestrating a larger model for coordination and a faster one for execution could make multi-model workflows more practical, if the reported performance holds up outside the company’s tests.
The release also gives users a concrete price and broad set of entry points, rather than limiting the model to a single developer product. Its long context and multimodal claims could matter for teams working across dense documents and recorded material, though the announcement does not establish how reliably it handles every such task.
Our read
This is a substantial product release, not a model-name reshuffle: Google is combining a fast general-purpose model with automatic model selection and reporting measurable trade-offs in speed and quota. The numbers are promising, especially for teams running repeated agent tasks, but they remain Google’s own results. Try Hydrogen on a representative workload and compare the bill, speed and output quality before handing it the keys to anything important. Benchmarks are a useful map; they are not the territory, as usual.
What to watch
- Whether independent tests reproduce the reported speed, benchmark and fabricated-claim improvements.
- How Hydrogen’s quality and cost compare with Carbon and rival models on real workloads.
- Whether Antigravity’s Orchestrator delivers its reported quota savings across a wider range of tasks.
- How well Hydrogen handles long-context multimodal work in day-to-day use.
Discussion spark: Would you let an orchestrator choose between faster and more capable models automatically, or do you want to control that trade-off task by task?
Sources and evidence
not affiliated with, endorsed by, or operated by Google or Google DeepMind