Google launches Gemini 4 Hydrogen and Orchestrator routing; model lands on Vercel AI Gateway

Google has released Gemini 4 Hydrogen, a new model for everyday coding and agentic work, alongside a model-selector option called Orchestrator that hands tasks back and forth between Hydrogen and the frontier Gemini 4 Carbon. The announcement landed on Google's Antigravity blog on November 12, and on the same day Vercel added Hydrogen to its AI Gateway, making it usable from popular coding agents outside Google's own tools.

The Hydrogen model

According to Google, Hydrogen stands on its own for routine feature work and debugging. It scores 51 on the Artificial Analysis Intelligence Index, one point ahead of Argon, and reaches 54 at its highest setting. On Terminal-Bench 4.1 it posts 57%, versus 18.2% for Gemini 3.8 Flash, the model Vercel describes as Hydrogen's predecessor. Reasoning depth is adjustable across four levels: low, medium, high (the default), and a new "xhigh" tier.

Vercel's changelog adds behavioral detail: Hydrogen sustains longer engineering tasks, turns out more thoroughly researched documents, spreadsheets and presentations in fewer agent steps, and keeps latency in the same class as 3.8 Flash. Vercel also credits it with tighter adherence to system instructions, formatting requirements and task boundaries; steadier tool orchestration across long conversations; stronger visual reasoning for multimodal work; and improved grounding, which Vercel says cuts down on unsupported claims in search-backed answers and fabricated tool or test results in coding reports.

Orchestrator and the routing problem

Google's stated motivation is that running a frontier reasoning model on every step of a routine task slows the developer loop and burns quota, while the industry's common fix — a static "model router" that infers task difficulty from the first prompt and commits the entire job to one model — tends to break down as agents run longer and their needs shift mid-task.

Orchestrator instead switches between Carbon and Hydrogen dynamically, in both directions, over the course of a single task. In Google's example, Carbon starts in a read-only exploration phase: mapping dependencies, clearing up ambiguities and drafting a plan with explicit triggers for replanning. Hydrogen then takes execution, applying edits, updating packages and running tests. If Hydrogen stalls, control returns to Carbon to inspect the state, repair the plan and resume. Google reports that on Terminal-Bench 4.1, Orchestrator reached task success close to Carbon's while consuming 24% less quota on average and finishing 19% faster.

For teams that want explicit control, the Antigravity SDK lets developers wire the two models together directly. Google's demo is an open-source issue-triage system in which Hydrogen agents handle triage and quick coding iterations while a Carbon agent takes on larger refactors, new features and security checks.

Hydrogen and Orchestrator arrive shortly after Gemini 4 Carbon, which Google launched earlier in the week together with /swarm for parallel, multi-worktree engineering tasks. According to the blog, tens of thousands of swarms were started on Antigravity within days.

Access and availability

Google says Hydrogen is rolling out to paid Google AI Pro and Ultra subscribers, while Orchestrator is limited to Google AI Ultra for now, with more models and plans promised later. Enterprise customers can reach both through Gemini for business.

On Vercel's side, the model is live on AI Gateway under the name google/gemini-4-hydrogen. Running npx vercel ai-gateway setup with the latest Vercel CLI wires it into agents including Claude Code, Codex and Cursor, and it can also be tried in the model playground. Vercel notes that AI Gateway provides a unified API with usage and cost tracking, retries, failover, zero data retention support, per-key budgets and routing rules, and it passes through provider pricing with no markup and no platform fee on inference, including on bring-your-own-key requests.

Why it matters

Orchestrator turns model routing from something developers configure into a feature the runtime handles itself, a direct answer to the cost and latency problems of long-running coding agents. The numbers Google cites — near-Carbon success rates at materially lower quota and higher speed — point toward systems that spend expensive reasoning only where it pays off. Just as notable is the distribution: same-day availability on Vercel's gateway means Hydrogen reaches developers building on non-Google agents immediately, rather than staying locked inside Google's own products.