Google announced a new AI model, "Gemini 4 Hydrogen," on November 12, 2026 (US time). It is now generally available through the Gemini API for developers and is also rolling out worldwide in Google Search. The successor to Gemini 3.8 Flash, it improves performance mainly in coding and tool-driven workflow automation. In a workflow automation evaluation run by Zapier, it achieved a success rate nearly identical to that of the higher-tier Gemini 4 Argon at roughly half the estimated cost. However, strengths differ by model in software bug fixing, terminal operation, and scientific tasks, and actual cost-effectiveness varies with the use case and reasoning settings.
What has improved over Gemini 3.8 Flash
According to the Google DeepMind model card, Gemini 4 Hydrogen was developed on the basis of the higher-tier Gemini 4 Carbon. Google positions Carbon and Argon for complex work that requires long, extended reasoning, and Hydrogen for everyday coding and tool-assisted task processing. When working with large codebases in particular, it can run reads and searches across multiple files in parallel and use the results to narrow down what to investigate and where to make fixes. This kind of processing matters for AI agents that proceed step by step while using multiple tools.
In the "Terminal-Bench 4.1" evaluation published by Google, Gemini 3.8 Flash had a success rate of 18.2%, while Hydrogen recorded 57.0%, an improvement of 38.8 points. Google also says that Hydrogen at its default reasoning level, High, completes tasks 2.5 times faster than the old Flash and generates content not grounded in fact less than one-fifth as often as the earlier model. However, these are results published by Google, and similar improvements are not guaranteed in real-world work.
When migrating from the old model, note that the default reasoning level has changed. According to Google Cloud's migration guide, the default for Gemini 3.8 Flash was Medium, but for Hydrogen it has been raised to High. If you use the API without specifying a reasoning level, switching models therefore also changes the reasoning setting. This could affect not only answer quality but also processing time, token consumption, and running costs.
Rivaling the higher-tier model in workflow automation at about half the cost
Zapier measured Hydrogen's performance in workflow automation in concrete terms. On November 10, the company published results from evaluating Hydrogen across different reasoning settings on 657 simulated tasks modeled on work such as customer management. The evaluation used "AutomationBench 1.0.6," and a task counts as successful only if the business data after execution meets all the conditions set for that task. Even if the AI generates a plausible explanation, it does not count as a success unless the required data updates and operations have been completed.
According to Zapier's benchmark report, Hydrogen at the High reasoning level completed 336 of the 657 tasks, a success rate of 51.1%. That is just 0.2 points behind the higher-tier Gemini 4 Argon, at about half the estimated cost.
| Gemini 4 model (High reasoning level) | Success rate | Estimated cost (USD per task) |
|---|---|---|
| Hydrogen | 51.1% | 0.82 |
| Argon | 51.3% | 1.63 |
| Carbon | 53.9% | 1.94 |
Hydrogen's estimated cost per task was $0.82, about 50% of Argon's $1.63 and about 42.3% of Carbon's $1.94. Carbon recorded the highest success rate at 53.9%, but its lead over Hydrogen is only 2.8 points. In these simulated tasks, at least, Hydrogen kept a success rate close to that of the higher-tier models while substantially cutting execution costs.
Note, however, that Zapier's costs are estimates that include cache usage, and the Hydrogen calculation used Gemini 3.8 Flash's list prices. Specifically, it assumes a cache read cost of $0.15 per million tokens, whereas Hydrogen's standard paid rate published by Google is $0.075. Costs calculated with the actual API prices may therefore differ from Zapier's estimates. The same report also recorded success rates of 44.7% for Claude Sonnet 5.5 and 51.4% for Claude Fable 5.2, but both models were run at the Max reasoning setting and with different cost-calculation conditions, so simple comparisons should be made with caution.
Raising the reasoning level does not always improve cost-effectiveness
With Hydrogen, success rate and running cost change substantially with the reasoning level. Zapier ran the same workflow evaluation at three levels, Low, Medium, and High, and calculated costs including failed attempts. It also divided the estimated cost of all attempts by the number of successes to compare the cost required per successful task.
| Hydrogen reasoning level | Success rate | Estimated cost (USD per task) | Estimated cost per success (USD) |
|---|---|---|---|
| Low | 30.6% | 0.18 | 0.58 |
| Medium | 46.7% | 0.41 | 0.88 |
| High | 51.1% | 0.82 | 1.61 |
Raising the level from Medium to High improved the success rate by 4.4 points, from 46.7% to 51.1%. Meanwhile, the estimated cost per task doubled from $0.41 to $0.82, and the cost per success rose from $0.88 to $1.61. Furthermore, in the "Operations" category, which covers business operations, the success rate was 65.0% at Medium and 64.0% at High, so raising the reasoning level did not improve results.
These results suggest that always choosing the highest reasoning level is not necessarily the best approach to workflow automation. For example, for routine tasks that are easy to retry after a failure, a lower reasoning level may keep costs down. For tasks where failure has a large impact, the added cost of a higher success rate may be worth paying. In actual deployments, evaluation should account not only for API pricing but also for the number of retries and the cost of human review and correction.
Strengths differ by model in coding, terminal operation, and science
Hydrogen does not outperform rival models on every evaluation. The model card published by Google DeepMind compares the performance of Hydrogen, Claude Sonnet 5.5, and GPT-6.1 Terra across multiple benchmarks. Hydrogen led by a wide margin on "SWE-sweep," which evaluates software bug fixing, while on terminal operation and scientific tasks some rival models came out ahead.
| Benchmark | Gemini 4 Hydrogen | Claude Sonnet 5.5 | GPT-6.1 Terra |
|---|---|---|---|
| SWE-sweep | 13.5% | 3.1% | 5.8% |
| Terminal-Bench 4.1 (adjusted for invalid solutions) | 57.0% | 63.9% | Not listed |
| Agents’ Last Exam v1.1 | 71.6% | 60.4% | 69.8% |
| AutomationBench-AA | 76.1% | 71.8% | 60.2% |
| OSWorld-2.1 (offline tasks) | 77.6% | 83.9% | 74.2% |
| Chartography | 66.6% | 61.6% | 63.2% |
| Terminal-Bench-Science 0.1 | 51.4% | 59.9% | 58.3% |
| HLE-Diamond | 58.8% | 53.3% | 47.3% |
* Bold indicates the highest score among the listed models on each benchmark.
Hydrogen scored 13.5% on SWE-sweep, ahead of Sonnet 5.5's 3.1% and Terra's 5.8%. It also took first place on AutomationBench-AA, which evaluates workflow automation, with 76.1%, and beat the comparison models on Agents’ Last Exam, Chartography, and HLE-Diamond. On the other hand, on Terminal-Bench 4.1, which evaluates terminal operation, Sonnet 5.5 scored 63.9%, above Hydrogen's 57.0%. Sonnet 5.5 also had the highest score on OSWorld-2.1, which involves on-screen operation, at 83.9%, and on Terminal-Bench-Science 0.1, which has the model solve scientific tasks in a terminal, both Sonnet 5.5 and Terra beat Hydrogen.
These differences show the importance of choosing a model according to the work being delegated to an AI agent. Identifying and fixing a bug in code requires different abilities from operating a terminal or screen and carrying out multiple steps. Real development work often includes both kinds of steps, so it is hard to judge suitability from the results of a single benchmark.
Note also that the 76.1% for AutomationBench-AA in DeepMind's comparison table and the 51.1% measured by Zapier are based on different evaluation conditions. Zapier used AutomationBench 1.0.6 and counted a task as successful only if all completion conditions set for it were met. Even though the names are similar, the evaluation methods differ, so the two success rates cannot be compared directly.
Ahead of the higher-tier model on a composite index, but processing speed differs
On the Artificial Analysis composite index announced by Google, Hydrogen scored 51 at its default reasoning level of High and 54 at the maximum, XHigh. That is above Argon's 50. However, in a separate comparison table for Carbon published by DeepMind, Carbon scores 59 and Argon 50 on version 5.0 of the same index. Hydrogen surpasses the higher-tier Argon on some composite evaluations, but a gap with Carbon remains.
Processing speed in artificial evaluation environments is also an important point of comparison. According to figures listed by Artificial Analysis, the time needed to complete the composite evaluation tasks at the Low reasoning level was 0.6 minutes for GPT-6 Luna, 1.1 minutes for GPT-6.1 Terra, and 1.3 minutes for Hydrogen. These are results with all three models measured at Low, and they do not show response times at Hydrogen's default of High. In actual deployments, you need to align the reasoning level used and compare both success rate and processing time.
API pricing and migration conditions: what to watch when moving from the old Flash
Gemini 4 Hydrogen's standard API pricing is set at the same level as the regular pricing of Gemini 3.8 Flash. However, because an introductory discount applies to the old Flash through the end of 2026, comparing at November 12 prices makes Hydrogen twice as expensive per unit for both input and output. Introductory pricing has also been set for the higher-tier Carbon and Argon, but it is scheduled to end on November 30.
Comparing the standard API pricing of the main models gives the following.
| Model | Input price (USD per million tokens) | Output price (USD per million tokens) | Price changes / long-input conditions |
|---|---|---|---|
| Gemini 4 Hydrogen | 1.50 | 7.50 | No surcharge by context length |
| Gemini 3.8 Flash | 0.75 | 3.75 | $1.50 / $7.50 from January 1, 2027 |
| Gemini 4 Carbon / Argon | 2.00 | 10.00 | $4.00 / $20.00 from December 1, 2026 |
| Claude Sonnet 5.5 | 2.00 | 10.00 | — |
| GPT-6.1 Terra | 1.00 | 5.00 | For requests over 272,000 input tokens, input is 2x and output 1.5x |
* Base unit prices as of November 12, 2026. Excludes costs from cache usage and tool calls such as search.
Hydrogen's input and output unit prices are 25% lower than Carbon's and Argon's prices during the introductory period, and will be 62.5% lower after the December 1 price revision. GPT-6.1 Terra's standard unit prices are lower than Hydrogen's, but for requests with more than 272,000 input tokens, 2x the normal rate applies to the entire input and 1.5x to the output. For uses that feed in large volumes of material, a simple comparison of base unit prices is therefore not enough to judge actual cost.
Even for the same task, billing differs because models consume different numbers of reasoning tokens and make different numbers of tool calls. When adopting Hydrogen, it is important to align input materials and task completion conditions and to compare not only success rate but also cost per run and cost per success. When migrating from the old Flash, running costs should be estimated based not on the current discounted price but on the regular rates from 2027 onward.
Context length of up to about 1.05 million tokens and API specifications
According to Google Cloud's published specifications, the maximum combined input and output Hydrogen can handle is 1,048,576 tokens. For ordinary generation, the output limit is set at 65,536 tokens, and this allowance covers not only the final answer but also internal reasoning and tool calls. To continue generating long texts or complex processing, v1beta1 must be used for both the Interactions API and the GenerateContent API. The overall output limit, including continued generation, is the total token allowance minus the input, so the more material you load, the less capacity remains for reasoning and output.
The reasoning level can be chosen from Low, Medium, High, and XHigh, with High as the default. Sending a request that still specifies thinking_budget or MINIMAL, used with older models, returns an HTTP 400 error. Applications migrating from Gemini 3.8 Flash therefore need to review their reasoning-related parameters in addition to changing the model name.
For caching, it supports implicit caching that automatically reuses the common portions of inputs, with a minimum eligible input length of 4,096 tokens. For uses that repeatedly rely on shared system instructions or large reference materials, caching may help keep costs down. However, because cache pricing and default reasoning settings differ between the old Flash and Hydrogen, cost estimates based on previous usage conditions need to be updated.
"Orchestrator": combining a higher-tier model with Hydrogen
Google has also introduced "Orchestrator" in its development environment "Antigravity," which lets users switch among multiple Gemini models within the same work session. For example, Carbon can handle planning and assigning complex work, while Hydrogen carries out the actual code changes and verification in parallel in isolated work environments. Rather than leaving every step to the highest-performance model, the aim is to reduce processing time and quota consumption by using different models according to the work.
In a test Google ran on Terminal-Bench 4.0, a setup using Orchestrator maintained a success rate close to that of Carbon alone while cutting Google quota consumption by an average of 24% and shortening task completion time by 19%. However, no specific success rates or monetary conversions of quota consumption have been published. The result suggests that assigning a relatively lightweight model to execution steps could improve the overall efficiency of AI agents. The actual effect will depend on which steps are left to the higher-tier model and which are routed to Hydrogen.
Worldwide rollout in Google Search, with automatic model selection in the Gemini app
Hydrogen is being introduced worldwide not only in the developer API but also in Google Search. According to Google, when the new search system receives a question containing multiple conditions, it breaks it into the points to be researched and runs the corresponding search queries in parallel. By leveraging Hydrogen's reasoning ability, it aims to grasp the intent of a question in more detail and search more widely for relevant information and links. Depending on the answer, images, comparison tables, and dedicated interactive tools are also displayed.
In this update, the improvements in language understanding apply to all languages supported by AI Mode in Google Search. Users of each language, including Japanese as well as English, are covered, so how answers to complex questions are structured and how information is gathered may change. However, Google has not provided specific measured figures on how much search result accuracy or traffic to external sites will improve.
Hydrogen is also used in the Gemini app, but the new model does not handle every question. On the free plan with automatic model selection enabled, Hydrogen may be chosen for questions that need deep reasoning, while Flash-Lite handles most general questions. The app routes questions to an appropriate model based on their content, even if the user does not specify one.
The introduction to Search is also a notable change for website operators. If Hydrogen broadens the points and related links covered by a search, sites that previously rarely appeared in results may have more chances to be found by readers. On the other hand, if the summaries, comparison tables, and images shown on the search page resolve users' questions on their own, there may be cases where visiting external sites becomes unnecessary. Whether traffic to external sites actually increases cannot be determined without continually monitoring link impressions, click-through rates, and visits.
Gemini 4 Hydrogen can be seen as a product that emphasizes running cost and processing efficiency in addition to improved model performance. In workflow automation it achieved a success rate close to that of higher-tier models at low cost, but on some evaluations in terminal operation and science, rival models are ahead, so it is not a universal choice. Going forward, attention will focus on how much it can improve cost-effectiveness in actual work, and on what changes its large-scale introduction in Google Search brings to users' search behavior and traffic to external sites.