Gemini 4 Hydrogen tiers and native Orchestrator mode arrive in Google Antigravity

What happened

Google's Antigravity IDE now offers the full Gemini 4 model series, according to a detailed hands-on post by developer Evgenii Zinner on dev.to. The rollout adds a new fast execution model called Gemini 4 Hydrogen, a fourth "xhigh" thinking tier exclusive to it, inline quota meters, and a native Orchestrator mode that automates a workflow developers had previously been forced to build by hand.

The details come from a single practitioner's report rather than official Google documentation, so specifics like internal model identifiers and pricing should be read in that light.

The quota problem this solves

When Gemini 4 Argon first reached Antigravity, the author writes, its strong architectural reasoning came at a heavy cost against the Pro-tier quota. Developers responded with a manual "Planner/Executor split": begin a prompt on Argon to map out edge cases, stop it, then flip the picker to Gemini 3.8 Flash to actually write the code. Others prompted Argon to behave as a coordinator and draft strict task briefs for subagents. Because remaining token usage was buried in the Settings menu, the author even built a Windows desktop widget, AgentPulse, just to track how much Argon quota was left before starting a large refactor.

How Orchestrator mode works

The headline change is a first-class picker entry, Orchestrator (Argon), internally identified as gemini-pro-orchestrator-g1-pro. According to the post, every message sent in this mode triggers a deterministic, two-phase plan-and-execute protocol inside one conversation thread.

In the first phase, the turn starts on the frontier planning model with write and file-editing tools disabled at the harness level. Argon can only read files, search the codebase, and run read-only observation commands such as builds and test suites. When it signals completion, it must produce an exhaustive step-by-step specification written for an executor that cannot ask follow-up questions: exact files, symbols and changes, precise verification commands with expected outputs, and a closing "re-plan if" checklist naming the signs that an assumption has failed.

In the second phase, the harness hands the same conversation history to Hydrogen, which sees every file Argon inspected plus the structured plan, then works through the edits, builds and test runs without drawing on the Argon quota. If Hydrogen hits a re-plan trigger, needs an architectural decision, or the user types a follow-up, execution pauses and Argon is woken with a hidden plan-review payload to inspect progress and adjust the remaining steps.

Hydrogen and the quota split

The post describes Hydrogen as the execution-optimized member of the elemental Gemini 4 lineup (Carbon, Argon, Hydrogen). Rather than training a small model from scratch, Google reportedly distilled agentic trajectory patterns, schema adherence and code-editing priors from the heavyweight frontier models into a streamlined architecture tuned for high tokens-per-second throughput. The author found it noticeably sharper at tool-calling than Gemini 3.8 Flash, which still carried 3.x-generation quirks on dense multi-file tool schemas.

One operational detail stands out: inspecting Antigravity's internal model registry, the author reports that Hydrogen shares the same fast-execution quota pool as Gemini 3.8 Flash, entirely separate from the Argon/Orchestrator pool. Using Hydrogen, whether directly or as the execution half of Orchestrator mode, therefore does not consume planning turns.

The xhigh thinking tier

While Argon and 3.8 Flash offer low, medium and high thinking levels, Hydrogen adds a fourth: xhigh. Citing independent evaluations such as the Artificial Analysis Intelligence Index, the post claims Hydrogen at xhigh reaches a score of 54, matching or edging out standard Argon configurations on complex algorithmic and debugging benchmarks while running at roughly a quarter of the API cost, at $1.50 per million input tokens and $7.50 per million output tokens. The author's practical takeaway: when Argon quota runs dry mid-afternoon, switching straight to Hydrogen at xhigh delivers near-frontier reasoning depth from the more generous Flash pool.

At the top of the hierarchy sits Gemini 4 Carbon, described as a heavyweight frontier model for long-horizon architecture, security auditing and research. It is currently gated to Google AI Ultra subscribers, and the author, on a standard Pro plan, has not tested it.

Why it matters

This release signals a shift in where orchestration logic lives: out of user prompting and into the IDE harness. Deterministically disabling write tools during planning, forcing structured plan contracts, and separating quota pools all encode hard-won practices that developers previously improvised. If the reported benchmarks hold up, a distilled model reaching frontier-adjacent scores at a quarter of the cost also changes the economics of everyday agent-driven coding, making aggressive model use viable without constant rationing.