Google Releases Lightweight Model Gemini 4 Hydrogen, Launching Three Major Models in Six Weeks
Beijing time November 13, Google released the new model Gemini 4 Hydrogen, the third model it has released within six weeks, with a focus on faster response, lower cost, and agent execution capabilities. Hydrogen is open to users worldwide, covering the Gemini app, Google Search, AI Studio, API, and enterprise services, and free users can also use it through the automatic model selection feature. Its API pricing is $1.5 per million input tokens and $7.5 per million output tokens; on the Terminal-Bench 4.1 test, its score rose from 18.2% to 57%. Google also launched Orchestrator in its AI coding platform Google Antigravity, supporting dynamic switching between different models within the same task.
Why read this
SelectedAI score 82
Google launched three models in six weeks. This article sorts out the positioning, pricing, and multi-model orchestration approach of the lightweight Hydrogen, which can be compared with its division of labor with the flagship model.
Topics
Body · AI translation
After releasing its flagship model the day before yesterday, Google released another new model, Gemini 4 Hydrogen, on November 13 Beijing time—the third new model Google has released in six weeks.
English text is being prepared.
Unlike the previous two models, Google's new model Hydrogen no longer pursues the upper limit of intelligence, but instead focuses on faster response speeds, lower usage costs, and more efficient AI agent execution capabilities.
According to Google's explanation, Carbon and Argon are mainly responsible for complex reasoning, long-term planning, and highly difficult software engineering, while Hydrogen is better at quickly writing code, calling tools, and handling everyday tasks that require continuous execution.
For ordinary users, the more important change is that everyone can finally experience Google's new model. Previously, Google mainly restricted access to its flagship models to premium paying users, which caused considerable dissatisfaction. This time, Hydrogen is being launched to users worldwide, covering products including the Gemini app, Google Search, AI Studio, the API, and enterprise services. Free users can also use Hydrogen through the automatic model selection feature in the Gemini app.
According to the intelligence index ranking of third-party AI evaluation organization Artificial Analysis, Hydrogen currently ranks seventh on the list, while the flagship model Carbon, updated the day before, ranks second.
How intelligent is Hydrogen? According to Artificial Analysis data, at the highest reasoning tier, Hydrogen's intelligence index reaches 54 points, slightly below the 55 points of the highest tiers of Claude Sonnet 5.5 and GPT-6.1 Sol, but as a lightweight model it is already close to the global top tier.
Another metric that has been widely discussed is that, at the default high tier, the more lightweight-positioned Hydrogen scores 51 points, even slightly higher than the previously high-end-positioned Argon (50 points). However, there is still a gap compared with the flagship model Carbon's 60 points.
Based on users' current brief evaluations, feedback on Google's new model is good. One user mentioned, "A computer problem that 3.8 Flash couldn't solve was solved by it right away—a dramatic improvement." He believes Google's model is now capable of "arm-wrestling" with Claude.
Another overseas user said, "For the first time I feel like I'm using a good Google model—it's fast, and for a $20 subscription the usage limits are quite generous." He mentioned that he might be the person who has criticized Gemini the most online, but Google is indeed improving. "If they keep this pace, I can even see Google surpassing Anthropic within the next 3 months—they have excellent engineers and almost unlimited resources."
A domestic user also called Hydrogen their favorite new model recently: "Its coding ability is not weak, its aesthetics are up to standard, its Chinese ability is outstanding, its copywriting ability is also strong, and it's blazing fast." He likened going from OpenAI's Sol model to Hydrogen as "a leap from a slow green-skinned train to flying."
Of course, these users' initial experiences are somewhat subjective, and whether the model can maintain its advantages across a broader range of real tasks still requires more verification.
Price and model cost-effectiveness have always been an important card Google plays. The API pricing for Hydrogen is $1.5 per million input tokens and $7.5 per million output tokens. Artificial Analysis estimates that at the highest xhigh tier, the average cost for Hydrogen to complete a standard evaluation task is about $3.16, only about one-third of Claude Sonnet 5.5's highest tier, though still about 2.5 times that of GPT-6.1 Sol's highest tier.
Hydrogen's more obvious progress is also reflected in execution efficiency. Google's published data shows that, compared with the previous generation Gemini 3.8 Flash, Hydrogen completes tasks at the default high tier 2.5 times faster than the former. In the Terminal-Bench 4.1 test, which examines AI agents' terminal programming capabilities, its score rose from 18.2% to 57%, and instances of fabricating false information decreased by more than 80%.
Artificial Analysis's evaluation also shows that Hydrogen performs outstandingly in visual reasoning and reducing hallucinations. In the AA-Omniscience test, which measures model knowledge accuracy and hallucination, its hallucination rate is about 8%, which Artificial Analysis noted is "one of the lowest levels ever recorded," while the highest tiers of Claude Sonnet 5.5 and GPT-6.1 Sol were 43% and 52%, respectively.
Google is also trying, from a product strategy perspective, to have different models divide the work. Along with the release of Hydrogen, Google also launched Orchestrator in its AI programming platform Google Antigravity, allowing users to dynamically switch between different models within the same task.
In its official blog, Google gave an example: in software development, Carbon and Argon can handle complex system architecture design, technical judgment, and analysis of difficult problems, while Hydrogen takes on higher-frequency execution work, including writing code, running tests, and verifying modifications.
According to Google's test results, in Terminal-Bench 4.0, this multi-model system achieved a task success rate close to Carbon's, while reducing average quota consumption by 24% and increasing task completion speed by 19%.
"It feels like Google is back—the 'big three' live up to their name, and the one lagging furthest behind is now OpenAI," one user lamented. The AI large-model industry's fortunes turn in cycles, and the pressure now shifts to OpenAI.