Gemini 4 Hydrogen, announced on November 12, 2026 by Koray Kavukcuoglu on the Google blog, was positioned as the speed and efficiency-oriented model of the Gemini 4 family; Google says it is making the model available globally for daily coding, agent-based workflows, the Gemini app, and Search.
According to Google, Hydrogen scores 51 on the Artificial Analysis Intelligence Index at the default high consideration level, is offered at $1.50 per million input tokens and $7.50 per million output tokens across the entire 1 million token context window. According to Abner Li on 9to5Google on November 12, 2026, Google did not use the Flash name in this model, and Hydrogen is seen as the replacement level for Gemini 3.8 Flash, which was released in September.
Image: Google
What is Gemini 4 Hydrogen positioned for?
Google states that it designed Gemini 4 Hydrogen for fast execution and wide use, not for the most difficult and long-running reasoning tasks. According to the company's positioning, Gemini 4 Carbon and Gemini 4 Argon stand in the upper class on the side of complex research, deep analysis and long-term reasoning. Hydrogen, on the other hand, aims to produce faster results in daily coding, debugging, tooling and multi-step agent workflows.
This distinction makes the product strategy of the Gemini 4 family more understandable. With the introduction of Gemini 4 Carbon, we saw that Google established a separate model layer for the most complex jobs. The subsequent launch of Carbon to Ultra subscribers showed that this model was placed in a more expensive and high-end usage area. Gemini 4 Argon is positioned as a separate option on the code and analysis side.
This is where Hydrogen's importance comes into play: Google wants to bring Gemini 4 to wider audiences' daily use, not just for high-end subscribers or heavy research work. This is why the company's statement to everyone is critical. This statement includes the fact that even free users can access Hydrogen with automatic model selection through the Gemini app at the appropriate prompts.
What does the Artificial Analysis score say?
According to the data shared by Google, Gemini 4 Hydrogen gets 51 points at the default high thinking level in the Artificial Analysis Intelligence Index. On the same scale, Gemini 4 Argon is at 50 points. Hydrogen reaches 54 points at the new xhigh thinking level.
These scores show that Hydrogen is not just described as a light and fast model. Google says the Hydrogen generally finishes faster than models in its class. However, there is an important limit here: Since the names and scores of the competing models are not given in the blog post, we cannot clearly see which models this comparison was made with.
The Artificial Analysis score alone should not constitute a purchasing or usage decision for the reader. Hydrogen's main claim is to make latency, cost, and token usage more predictable while keeping quality at a certain level. Therefore, the value of the model will be measured by how quickly and consistently it behaves in the real workflow, rather than a single score.
How much does Gemini 4 Hydrogen cost?
According to the Google AI Studio pricing page, Gemini 4 Hydrogen is available at a fixed price across the entire 1 million token context window. For paid usage, the price is $1.50 per million input tokens and $7.50 per million output tokens. Thought tokens are also included in the output price.
In the free tier, input and output appear to be free. However, it is stated that the data used in the free tier can be used for product development purposes. In the paid tier, data is not used in product development. This distinction is as important as price, especially for institutions and privacy-sensitive developers.
The Google AI Studio price page includes $0.075 per million tokens for context caching only within the scope of automatic caching. On the grounding side with Google Search, 5,000 requests per month are free. This right is shared with Gemini 3 and later models. Once the free limit is exceeded the price is $14 per 1,000 requests.
How does Orchestrator use Carbon and Hydrogen?
In real software development processes, Google highlights an order in which multiple models work together instead of a single model doing everything. In this scheme, metamodels such as Carbon and Argon are used for architectural decisions, hard errors, and complex reasoning. Hydrogen, on the other hand, takes part in rapid execution steps such as writing code, running tests and verifying changes.
The new Orchestrator in Google Antigravity can dynamically switch between models such as Carbon and Hydrogen in the same session. According to the company's Terminal-Bench 4.0 data, Orchestrator provides task success close to Carbon, uses an average of 24% less quota and completes jobs 19% faster.
In parallel agent scenarios, Carbon can take on the role of coordinator. Hydrogen, on the other hand, can make and test changes simultaneously as workers and validating agents in separate workspaces. This approach shows that Hydrogen should be considered as a highly efficient execution engine rather than a stand-alone model.
What has changed compared to Gemini 3.8 Flash?
According to information published on 9to5Google on November 12, 2026, Google did not use the Flash name for Hydrogen. The model selector in the Gemini application includes 3.5 Flash-Lite for quick answers, 4 Hydrogen for daily help, 3.1 Pro for detailed research, 4 Argon for code and analysis, and 4 Carbon for the most complex jobs.
According to Google's data, Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash at the default high setting. In Terminal-Bench 4.1, Hydrogen gets 57% result, while 3.8 Flash remains at 18.2%. Google describes this difference as more than a threefold improvement.
The company also says Hydrogen reduces fraudulent claims by more than 5x. It is stated that the model adapts more smoothly when the vehicle call fails or the call returns empty. In early testing from partners, Hydrogen runs 2 to 2.8 times faster than 3.8 Flash and reduces token usage by 12 to 23% in complex graph, SQL, and data-intensive enterprise workflows.
Which usage areas stand out?
Google specifically highlights Hydrogen for on-repository coding and debugging. According to the company, the model makes targeted and surgical edits and does not rewrite irrelevant files. It is stated that it surpasses models of similar levels in SWE-sweep, but the figure for this test is not shared in the blog post. Making file reading and searching calls in parallel is also offered as a speed advantage in the developer workflow.
The picture shared by Google in open-ended agent missions is more ambitious. Hydrogen is ranked 5th in arena.ai's live benchmark Agent Arena. It is stated that it is the leader in its class with 58.8% in HLE-Diamond, ranks 1st overall in Vals Tax Agent Bench with 52.3%, and shares 1st place in Vals Legal Research Bench with 55.8%. It is said to be the leader in its class in AutomationBench-AA, but no figures are given.
Multi-modal information business is also one of Hydrogen's important areas. The model can create reports sourced from PDF, tables, graphics, audio and video recordings with local input of 1 million tokens and output of up to 1 million tokens. It is stated that it surpasses other models with 53.4% in the new Multimodal Knowledge Agent benchmark and 71.6% in Agents' Last Exam v1.1.
On the web development side, Google highlights examples of live interface prototyping and moving from design to code. Case studies include Geminisland adding structure to a common voxel island with a single command, an e-commerce storefront with WebGL2 shaders, turning Google Sheets into an animated pixel canvas, and producing a Google Slides presentation from a YouTube keynote. It is also stated that with Gemini in Google Spreadsheets, interactive models from financial data are output more than 30% faster on the Sheets Canvas side compared to previous models.
What do the four levels of thinking do?
Gemini 4 Hydrogen; It offers four thinking levels: low, medium, high and the new xhigh. The default setting is high. According to Google, this structure provides more predictable scaling between quality, latency and cost. Users can only wait and pay for the required depth of reasoning.
The company says the Hydrogen is on par or better than some competing models in its class on all levels, generally running faster and using fewer output tokens. However, since the names of these models, their detailed scores at each level and price comparisons are not given in the blog post, it is difficult to independently evaluate the scope of the claim.
What's changing in the Gemini app?
In the Gemini app, Hydrogen follows instructions better, calls the right tools, and delivers reliable output with less back-and-forth. Google says the model remains more faithful to tone, format and conventions. A usable result is aimed at the first try for daily tasks such as email drafting, handouts and rewriting.
The model can create tables for product comparisons. It can pull context from uploaded files, Gmail and Google Drive, as well as connected apps like Airtable and Dropbox. Case studies include interactive 3D sphere and music engine from real-time lightning data, WebGL2 water and glass visualizer, and archival video editing about the element Hydrogen.
According to a post on the Gemini app X account dated November 12, 2026, free users will access Hydrogen with automatic model selection when available for the prompt. The model is being rolled out globally to all users across desktop, mobile and web.
What will Hydrogen change in Google Search?
According to information shared by Robby Stein on the Google Search blog on November 12, 2026, Hydrogen is described as the biggest model upgrade for Search to date. This upgrade is available globally to all users.
Google states that with Hydrogen, multi-part, spelling incorrect or very specific questions will be better understood. Search results will feature more visual answers, images, structured answers, comparison tables and special tools. By strengthening the query spectrum technique, it is aimed to include a wider range of links and previously undiscovered sites in the results.
This section is also important for publishers. If Google really showed a wider range of links, well-structured and well-sourced content could gain more visibility. However, it will be necessary to see the practical effect of this with Search Console data.
What do security claims cover?
Google says Hydrogen reduces overshooting in agent missions by up to 81% and underserved service by up to 79% compared to 3.8 Flash. It is stated that in tool usage and coding tests, misaligned actions and task shortcuts decreased by 12 times.
It is explained that the probability of exceeding permission limits in impossible or disabled tasks is up to 2 times lower. In the in-house agent honesty test, a 96.7% honesty rate is shared. Over 200,000 audited rounds report 99.9% train-of-thought readability and zero deceptive reasoning edits.
On the indirect demand injection side, Hydrogen is stated to be more than 2.3 times durable in the Gray Swan IPI benchmark. The false rejection rate in HealthBench Pro is said to remain below 0.4%, and the model is said to better distinguish between biosecurity risks and legitimate medical and scientific questions. Google states that the details are included in the model card.
Who can use Gemini 4 Hydrogen?
Gemini 4 Hydrogen opens globally starting November 12, 2026. Developers can use the model in Google AI Studio via the Gemini API and in the paid tier of Android Studio. Paid Google AI Pro and Ultra subscribers have access to Hydrogen within Google Antigravity.
Access for institutions is provided via API or model menu within the scope of Gemini for business. Google Antigravity and Gemini in Google Sheets for eligible business subscriptions are also available for enterprise access. Access for everyone is offered through the Gemini application model menu and Google Search. Gemini in Google Sheets is rolling out to Google AI Pro and Ultra subscribers first. Google DeepMind's X post dated November 12, 2026 also counts Gemini in Chrome for AI Pro and Ultra subscribers.
What do the benchmark and price summary look like?
The Hydrogen result shared by AlanGoogle is given asNotArtificial Analysis Intelligence IndexHigh: 51, xhigh: 54Argon 50Terminal-Bench 4.1%573.8 Flash: 18.2%Agent Arena5. Sıraarena.ai live benchmarkHLE-Diamond%58.8Leader claim in its classVals Tax Agent Bench52.3%Overall 1st place claimVals Legal Research Bench55.81%. shares the rankMultimodal Knowledge Agent%53.4New benchmarkAgents' Last Exam v1.1%71.6It is stated that it surpasses other models
UsagePrice or accessDetailGemini API paid input$1.50 per 1 million tokens1 million tokens fixed in context windowGemini API paid output$7.50 per 1 million tokensThinking tokens includedFree layerInput and output freeData can be used in product developmentContext cache$0.075 per 1 million tokensAutomatic caching onlyReasoning with search5,000 requests per month freeAfter $14 per 1,000 requestsDevelopersGoogle AI Studio, Gemini API, Android Studio paid tierGlobal as of November 12, 2026EveryoneGemini app and Google SearchAutomatic model selection on eligible prompts
Review: What does Hydrogen mean?
What Gemini 4 Hydrogen really means is Google's effort to reduce high-end AI models from a single expensive research tool to an everyday production line. Carbon and Argon are positioned for more difficult reasoning tasks, while Hydrogen handles the fast, frequently repeated part of the workflow. This is a more realistic usage model, especially for software development, data processing, file-based reporting, and in-house agent processes.
How to look at the numbers?
If Google's claims are true, the Hydrogen could be not only faster but also more controlled and less wasteful than the Gemini 3.8 Flash. 2.5x task speedup, 57% results in Terminal-Bench 4.1, more than 5x reduction in fabricated claims, and 12% to 23% lower token usage across partners are significant from a developer cost perspective. However, it is important to remember that most of these numbers come from Google's own tests and are not independently verified.
On the price side, Hydrogen is in a different league from Carbon, at $1.50 per million input tokens and $7.50 per million output tokens. However, there is no direct price comparison with Gemini 3.8 Flash on the Google blog. Therefore, the real cost advantage of Hydrogen over the old Flash layer can only be understood through independent trials on the same workloads.
It is also necessary to be careful on the benchmark side. The blog post does not give the names and scores of competing models, but only refers to the models in their class. Additionally, some tests, such as Multimodal Knowledge Agent, are newly announced or less well-known.
