JOURNAL / 2026.11.13

Gemini 4 Hydrogen and Orchestrator: switching models while an agent works

Google broadens access to Gemini 4 and proposes alternating planning and execution between models. The opportunity is to complete more work within resource limits; testing it also means evaluating the handoffs.

Google introduced Gemini 4 Hydrogen on November 12, rolling it out across the API, AI Studio, the Gemini app and Search, among other channels. It targets fast execution and everyday agents. Announced pricing is $1.50 per million input tokens and $7.50 per million output tokens, with no surcharge for context length.

The more interesting proposal comes from combining it with other models. Orchestrator in Antigravity can move from Carbon to Hydrogen within one task and back again if execution runs into difficulty. The official example assigns exploration and planning to Carbon, and edits and tests to Hydrogen. Choosing a model at the start and alternating during work are different decisions: difficulty can emerge after opening the repository.

Imagine a dependency upgrade. Changing versions and running tests may be routine; discovering that a library changes date semantics calls for revisiting the plan. My reading is that a handoff has value if it recognizes that distinction in time. The hard question becomes what the system preserves when switching models: the reason behind each decision, errors already ruled out, and conditions requiring it to stop.

The execution model's improvement also has concrete evidence. The model card reports 57% on Terminal-Bench 4.1, an evaluation of agents in a terminal; the announcement places Gemini 3.8 Flash at 18.2%. The published methodology specifies high effort and a mini-swe-agent vision harness for Hydrogen, averaging five attempts per task. That does not turn five opportunities into a single infallible run. Nor does it by itself measure migration quality in a repository outside the benchmark.

For the combined system, a discrepancy argues against relying on the advertised savings: the general blog attributes Orchestrator's results to Terminal-Bench 4.0, while Antigravity's blog names 4.1. Both describe success close to Carbon with lower quota consumption and elapsed time. Until the version and configuration are clarified, I would treat that comparison as a vendor indication. Product quota, billed cost and engineering hours are different measures.

For integration, another change is less conspicuous. The technical documentation specifies a shared budget of 1,048,576 tokens for input and output, including reasoning. Long outputs require continuation after segments of up to 65,536 tokens. Reading “one million input and one million output” as two independent allowances would produce a faulty budget. The developer guide also warns that custom temperature values are ignored and explicit caching is unsupported. A migration needs to review these conventions alongside the model identifier.

Hydrogen's model card describes it as an adaptation of Carbon. It acknowledges hallucinations, possible delays and increased token consumption at higher reasoning settings. The full safety report is still promised for the coming days; published automated evaluations record more unjustified refusals than 3.8 Flash. This documentation identifies outstanding questions, but cannot yet settle an assessment of production behavior.

I would test three paths on the same assignments: Hydrogen alone, the planning model alone, and the combination. I would count accepted deliveries, full cost, time and human review, and examine cases where a handoff repeats an investigation or loses a constraint. This is my proposal, not an experiment we have performed. The combination deserves adoption if it improves completed work: adding a second model also adds a conversation that can go wrong.

What persuades me about this release is that it makes it possible to discuss where to spend reasoning within a process. What I would still want demonstrated is whether the system recognizes when a task stops being routine. That will determine whether switching becomes a useful engineering tool or simply makes execution harder to explain.

Primary sources consulted on November 13, 2026. We have not run Hydrogen or reproduced the evaluations. The examples, interpretation of the release and proposed comparison procedure are our own analysis.

← Back to journal