Google’s Gemini 4 Hydrogen is cheaper. The real test is finished work.

Google released Gemini 4 Hydrogen on November 12, 2026, offering a lower-priced route into its latest model family without the restricted opening given to Gemini 4 Argon and Carbon. Access reaches developers and free Gemini app users from launch. The immediate change is broader availability at lower quoted token rates; the evidence for cheaper, successfully completed work is still to come.⁠1⁠2

Those rates compare with Argon’s introductory prices of $2 per million input tokens and $10 per million output tokens. SiliconANGLE reports Hydrogen’s prices, and listings from OpenRouter and benchmark operator Artificial Analysis corroborate them. A Hydrogen-specific row in Google’s own pricing table was not verified in the available evidence, so the distinction between reported rates and primary pricing confirmation remains important.⁠1⁠14⁠16

Tokens are the units used to meter what goes into a model and what comes out. These are therefore usage prices, not a fixed fee for finishing a task. A model that needs more attempts—or consumes more tokens on each attempt—can erode a lower unit price. Comparing the same quantities of input and output is straightforward; comparing the cost of a usable result is harder.

The reported rates cover a context window of roughly one million tokens: the material a model can work with at once. The listings do not use identical figures. SiliconANGLE describes one million, OpenRouter lists 1.0 million, and eesel AI gives 1.05 million. That is not evidence of separate products, but the exact maximum is not settled by the primary specifications available here.⁠1⁠14⁠19

The most attention-grabbing comparison is close. Google says Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default “high” thinking setting, against Argon’s 50. A new fourth thinking level, “xhigh,” raises Hydrogen’s reported score to 54. The higher setting is a different configuration, not a substitute for the default-setting comparison.⁠1

Artificial Analysis’s located model listing also records Hydrogen at xhigh with a score of 54, although the date that result was posted or updated was not established. Equivalent direct operator records were not verified for Hydrogen’s default-setting 51 or Argon’s 50; those remain Google’s reported comparison. An index score is not the percentage of a developer’s tasks the model will complete correctly.⁠16⁠1

The distinction matters because “cheaper and smarter” compresses several choices into one slogan. A developer must choose a thinking setting, measure the tokens used and test whether the result works. A narrow lead on a general index is a reason to evaluate Hydrogen, not a reason to assume it will replace Argon on every job.

Google DeepMind senior vice president and Google chief AI architect Koray Kavukcuoglu describes Hydrogen as “built for rapid execution.” The accompanying speed claim uses a different comparator: Gemini 3.8 Flash, released in September, rather than Argon. Google says Hydrogen finishes tasks 2.5 times faster at its default setting. That is a vendor-reported test result, not a demonstrated speedup across every workload.⁠1⁠2

Google also reports a substantial gain over Flash on Terminal-Bench 4.1, which tests work performed at a command line. Hydrogen scored 57%, against Flash’s 18.2%—more than triple the older model’s result. It is a notable improvement on that test, but it does not establish the same improvement in a customer’s software or workflow.⁠1

Alongside Hydrogen, Google introduced Orchestrator in Antigravity, its coding tool. Rather than relying on one model throughout a coding session, the feature chooses between Hydrogen and the more capable Gemini 4 Carbon as work progresses. The proposition is selective use of the larger model: reserve its resources for the parts that need them.⁠1⁠3

In Google’s reported testing, Orchestrator approached Carbon’s success rate while using an average of 24% less quota. “Approached” is doing necessary work in that claim: the result was not described as an exact match. Nor does a 24% reduction in quota establish an equivalent reduction in a user’s bill. It is evidence for a resource-saving approach under Google’s test conditions, not a universal savings promise.⁠1

The early testimonials put a more recognisable problem behind those numbers. Andrew Carr, a former OpenAI scientist who is now chief scientist at 3D animation startup Cartwheel, said Hydrogen caught two problems in an approach his team had been stuck on for 10 days. Discovery Loop, co-founded by former Google chief scientist Jeff Dean, was another named early tester. These are launch accounts selected by Google, not controlled evidence of how often another team will get the same benefit.⁠1

Hydrogen’s distribution is less conditional than its performance pitch. Google’s November 12 Search announcement explicitly says the model is available globally in Search for all users. SiliconANGLE also reports launch access through the Gemini API, Google AI Studio and the free Gemini app, with Google AI Pro and Ultra subscribers able to use Hydrogen in Antigravity. No numerical allowance for free app users is established here; broad access should not be read as unlimited use.⁠8⁠1

That opening contrasts with the rollout of its larger siblings. Argon first went to vetted cybersecurity defenders in Google’s Fairwind Program, then reached paying developers on October 13 after nearly two weeks of restricted access. Carbon also began with Fairwind members before opening to paying developers and Google AI Ultra subscribers on November 10. Hydrogen followed two days later, bringing the family to three releases in just over six weeks. The different rollout does not establish identical capabilities or safeguards.⁠1

The next dated commercial checkpoint is November 30, when Argon’s introductory pricing is scheduled to end. SiliconANGLE reports subsequent rates of $4 per million input tokens and $20 per million output tokens. If that transition takes effect as described, Hydrogen’s listed price advantage will widen considerably. The future rates, like the current comparison, still need confirmation against Google’s pricing record.⁠1

Hydrogen gives more people a cheaper model to try now. Whether it becomes the cheaper model to rely on depends on less glamorous measurements than a launch-day ranking: correct results, retries, thinking settings and total resource use. The useful comparison is not simply the price of a million tokens. It is the price of work that does not have to be done again.