Google brings a new Gemini model to Search
Google says Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash. Developers get four reasoning settings.
By Saeed Ezzati7 min read
The audio edition
Listen to this newsletter
0:005:18
Read transcript
Google is putting its speed-focused Gemini 4 Hydrogen model into Search, the Gemini app, developer tools and business products, with a clear pitch: get useful work done faster, without giving up too much reasoning depth. The global rollout began November 12. In Search, Google says Hydrogen is better at handling multipart questions, typos and specific requests, and can produce more visual responses while surfacing a broader range of relevant links. The performance figures come from Google, not an independent test. The company says Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash at its default setting, and cuts fabricated claims by more than fivefold. On Terminal-Bench 4.1, Google reports a score of 57 percent, compared with 18.2 percent for Flash. Those benchmarks make a case for the model; they don’t guarantee the same gains on every real-world task. For developers, the Gemini API costs $1.50 per million input tokens and $7.50 per million output tokens, with flat pricing across its one-million-token context window. Four reasoning settings—low, medium, high and xhigh—let teams trade speed and cost against deeper reasoning; high is the default. Access includes AI Studio and the paid tier of Android Studio. Hydrogen is also designed to work alongside Google’s other models. In Antigravity, a tool for coding, Orchestrator can switch between models such as Carbon and Hydrogen during a session. In another setup, Carbon coordinates while Hydrogen agents build and check changes in parallel, in isolated workspaces. Hydrogen can also code and debug on its own. So the rollout is not simply a replacement story: Google is offering a faster model for standalone tasks and as a worker inside a larger system. That push to make AI systems more useful also raises a practical question: how much engineering does it take to move a workload onto different hardware? Anthropic says its Mythos model helped get one Claude model’s prompt processing running on AMD MI355X GPUs in roughly three days, compared with six to twelve months using traditional engineering. Anthropic says Mythos wrote or adapted nearly all the chip-specific code, with almost no human intervention. The claim comes from the company’s investor materials and has not been independently verified. The scope matters: this was one model’s prompt-processing workload, not a rapid migration of every Claude workload or Anthropic’s entire operation. The example points to a possible reduction in the software costs that make switching away from Nvidia difficult. Anthropic says it now uses Amazon Trainium, Google TPUs, Nvidia GPUs and AMD GPUs, and describes their capacity as an interchangeable pool. Faster porting could make that flexibility more practical. The thing to watch is whether similar results hold across more models and workloads—not just whether one chip-specific port can be completed quickly. And as AI takes on more work, the other test is whether systems can resist misuse without refusing too much legitimate help. OpenAI says training models on jailbreak prompts generated by an AI attacker improved resistance to both fixed attacks and attackers that adapt during a conversation. The controlled study covered cybersecurity, biology and violent activity. Researchers used successful adversarial prompts as reinforcement-learning training data, then tested the models against different attacks. OpenAI reports stronger robustness than its baseline in both fixed and adaptive evaluations. That suggests the method may generalize beyond the specific prompts used for training, but the finding is limited to the study’s evaluated attacks. Separately, OpenAI reports that GPT-6 Astra answered safely on 91.5 percent of fixed cybersecurity jailbreaks, versus 59 percent for GPT-5.6 Sol. Those figures measure the models’ responses to evaluated prompts, not the complete safeguards used in production. OpenAI also says Astra’s tests showed fewer mistaken refusals of legitimate requests—a useful balance to track alongside attack resistance. That gap between a measured result and the larger conclusion matters outside AI benchmarks, too. WIRED reports that an internal Nvidia suggestion-box post urging the board to consider employees’ views on Jensen Huang’s departure drew more than 1,600 comments, many calling for his return. That is a comment count, not a count of individual employees seeking reinstatement, and it is not a companywide vote. Employees also asked why Huang left before the board’s investigation was complete. WIRED reports that some had expected the investigation to finish before a major leadership decision. The board named finance chief Colette Kress interim CEO and is searching for a permanent successor. Huang’s departure followed reports alleging an affair with a direct report; WIRED also reported his admission of infidelity, which is distinct from that allegation. Across these stories, the useful thing to watch is what the claims look like at operating scale: whether faster models deliver dependable gains on real tasks, whether chip portability extends beyond one workload, whether safety holds against new attacks, and how Nvidia’s board explains its decision. In each case, the headline is a starting point; the scope and evidence determine what it means in practice.
