Google released Gemini 4 Hydrogen on November 12, a fast model for coding and agents that scores 51 on the Artificial Analysis Intelligence Index, one point above its larger sibling Gemini 4 Argon. Input runs $1.50 per million tokens. Output costs $7.50 for every million. Google charges the same rates across the full 1 million-token context window. A fast execution tier that edges past a larger sibling on an index is the pitch for teams that run agents all day.
Google’s announcement, written by Koray Kavukcuoglu, positions Hydrogen below Gemini 4 Carbon and Argon, the two frontier models built for long-horizon reasoning. Hydrogen is meant for execution, the code-test-verify loop. On the default high thinking setting it scores 51 against Argon’s 50, and the new xhigh setting lifts it to 54. The index is a third-party measure, though the announcement does not include a full independent evaluation of the speed claims.
The more interesting product is Orchestrator, a feature in Google Antigravity that switches between models inside one session. Carbon handles the judgment calls while Hydrogen does the repetitive work. On Terminal-Bench 4.0, Google says Orchestrator came close to Carbon’s task success rates while using 24% less quota and finishing 19% faster. For parallel setups, Carbon coordinates and Hydrogen runs worker and verifier agents across isolated workspaces.
Against Gemini 3.8 Flash, the model it replaces for many teams, the gains are large on Google’s own numbers. On the default setting Hydrogen is 2.5 times quicker to finish, and it scores 57% on Terminal-Bench 4.1, where Flash scored 18.2%. Fabricated claims fall by more than 5x. Partners who tried it on SQL, graph, and other enterprise data jobs saw runs 2x to 2.8x faster, with token use down 12% to 23%.
Outside coding, Google cites several rankings. Hydrogen sits at #5 on Agent Arena, the live benchmark run by arena.ai. It leads its class on HLE-Diamond at 58.8% and ranks first on the Vals Tax Agent Bench at 52.3%. It ties for first on the Vals Legal Research Bench at 55.8%. These are vendor-reported figures, and the comparison sets are described only as “models in its class.”
Developers pick from low, medium, high, or xhigh reasoning depth. Higher levels cost more tokens and time, which makes the pricing question concrete. At $7.50 for each million tokens generated, an agent loop that burns millions of tokens on retries is a line item, so output-token efficiency matters as much as the sticker rate. Google says Hydrogen often uses fewer output tokens than comparable models, and that is the claim to test.
Safety figures arrive with the launch too. Google reports that, against 3.8 Flash, Hydrogen reduces overreach by as much as 81% on agent tasks and under-service by as much as 79%. It also reports a 96.7% score on its internal agent honesty evaluation. Those evaluations are Google’s own, and the company points readers to the model card for detail.
Availability is broad. Hydrogen is in Google AI Studio, the Gemini API, Android Studio, and Gemini for business. It is also in the Gemini app, Google Search for all users, and Gemini in Google Sheets. Paid Google AI Pro and Ultra subscribers can use it in Antigravity.
Teams on Gemini 3.8 Flash can run their own agent traces through Hydrogen this week at the default high setting, and measure total cost per completed task rather than per-token price. If the 2.5x speedup holds on real repositories, the Flash tier stops being the default choice for coding agents.
Google, “Introducing Gemini 4 Hydrogen,” blog.google, November 12, 2026.
