Google released Gemini 4 Hydrogen on November 12, making it available to all free users for the first time on launch day. It replaced 3.8 Flash with substantial gains in coding and intelligence benchmarks, but doubled the output price, while its reasoning setting consumes even more tokens. Model updates and usage costs are rising together.
What this is
Hydrogen succeeds Gemini 3.8 Flash and targets everyday coding and Agent workflows: Agent refers to a large-model application that autonomously breaks tasks into subtasks and invokes tools to execute them. It natively supports a 1-million-token context window. Google says the default setting runs 2.5 times faster, while its score on the Cursor coding leaderboard rose from 39.6% to 54.3%.
This time, Google did not roll out access in stages. AI Studio, the API, the Gemini App, Search, and developer tools all became available on launch day, including for free users. An Orchestrator mode also lets the planning model, Carbon, hand tasks back and forth with the execution model, Hydrogen; Google says this can reduce quota consumption by 24%.
Industry view
We believe Hydrogen's real impact is not a few additional leaderboard points, but that Google is starting to retire the low-cost Flash line outright. Gemini 3.8 Flash costs $0.75 per million input tokens and $3.75 per million output tokens, whereas Hydrogen's output price has risen to $7.50. For enterprises making high-frequency model calls, the reasoning setting may matter more than the sticker price: independent evaluations found that the xhigh setting generates an average of 304 million tokens, more than four times the leaderboard median.
The risks are equally clear. Benchmarks do not agree on which model is stronger: Hydrogen leads Sonnet 5.5 on SWE-sweep but trails it on Terminal-Bench. Reliability also remains unproven over the long term when a model autonomously decides which planning and execution models to use. Quietly replacing older models, introducing complex naming, and deeper inference consumption may make migration and budget management more difficult for developers.
Impact on regular people
For enterprise IT: The procurement threshold is shifting from whether a model can be called to whether calling it is cost-effective. Enterprises need to select model settings by task and set budget ceilings for inference tokens, tool calls, and automatic routing.
For individual professionals: Free users gained stronger coding and long-task capabilities on launch day, but model names and prices are now changing faster. Workflows dependent on a single product should retain alternatives to avoid sudden cost increases after an upgrade.
For the consumer market: Search and multiple applications are gaining stronger reasoning capabilities, but responses may become more complex and slower. Consumers will encounter more AI-generated content, making source verification and quality control increasingly important.