Introducing Gemini 4 Hydrogen

At its default high thinking level, Hydrogen scores 51 on the Artificial Analysis Intelligence Index versus Argon's 50, rising to 54 at the highest xhigh setting, while the company says it often completes tasks faster than models in its class. Pricing is flat across the full 1 million-token context window: $1.50 per million input tokens and $7.50 per million output tokens. For teams still on Gemini 3.8 Flash, Google positions it as a standalone upgrade — tasks finish 2.5x faster at the high default setting, Terminal-Bench 4.1 scores more than triple (57% versus 18.2%), and fabricated claims are cut by more than 5x, with cleaner handling of failed tool calls and empty search results.

The pitch extends to multi-model setups. Google is launching Orchestrator in Google Antigravity, which switches between models like Carbon and Hydrogen within one session; on Terminal-Bench 4.0 it reached near-Carbon task success rates while using 24% less quota on average and finishing 19% faster. Carbon can act as coordinator while Hydrogen powers worker and verifier agents across isolated workspaces. On coding benchmarks, Google says Hydrogen outperforms similar-tier models on SWE-sweep, and that in early partner testing across complex graphs, SQL, and data-heavy enterprise workflows it ran 2x to 2.8x faster while cutting token use 12% to 23% versus 3.8 Flash.

Benchmark claims are broad: #5 on Agent Arena for multi-step agentic tasks, 58.8% on HLE-Diamond, #1 on Vals Tax Agent Bench at 52.3%, a tie for #1 on the Vals Legal Research Bench at 55.8%, plus leadership on AutomationBench-AA, 53.4% on the new Multimodal Knowledge Agent benchmark, and 71.6% on Agents' Last Exam v1.1. It pairs native 1 million-token input with up to 1 million output tokens for synthesizing PDFs, spreadsheets, charts, audio, and video into cited reports. Demonstrations include Geminisland, a shared voxel island built from single prompts, and an e-commerce storefront produced in one shot with a WebGL2 topographic shader and live configurator, using the new Gemini front end design skill.

Open questions center on how the tiering holds up in practice: whether the near-Carbon results from Orchestrator on Terminal-Bench 4.0 generalize beyond that benchmark, how flat-context pricing and quota savings change total cost once parallel worker and verifier agents are in the mix, and whether enterprise teams split architect-versus-executor roles permanently or consolidate as frontier models get faster.