A coding agent has a clean plan for a routine change. A faster model could implement it. Then a test reveals that the plan was built around an assumption that no longer holds. The decision to spend more reasoning becomes obvious after execution has already started.
Google’s November 12 Hydrogen announcement and Antigravity Orchestrator rollout address this timing problem. Hydrogen is positioned as an efficient executor, while Carbon and Argon handle more difficult reasoning. Orchestrator can switch models during a session instead of committing to one at intake. The real question is whether a switch preserves enough state and quality to repay its cost.
Two benchmark claims with different meanings
Google reports that standalone Hydrogen achieves 57% on Terminal-Bench 4.1 versus 18.2% for Gemini 3.8 Flash in the comparison it presents. Separately, Google reports near-Carbon success with 24% less quota and 19% lower elapsed time on average for Orchestrator. The second result is about a product routing configuration. It does not estimate the accuracy or cost of every model switch.
There is a version discrepancy in the first-party write-ups: the Hydrogen launch associates the Orchestrator comparison with Terminal-Bench 4.0; Antigravity’s own article calls it 4.1. I could not resolve that attribution from the published descriptions. Use the figures as a vendor signal and record the exact suite, harness, and task set before comparing them with another benchmark run. Also, quota is a product unit. It cannot be converted directly to API dollars.
The launch quotes Hydrogen rates of $1.50 per million input tokens and $7.50 per million output tokens across the stated one-million-token context window. Those are launch API rates; the product’s rollout is a different matter. Antigravity lists Carbon and Argon Orchestrator options for Ultra and the Argon option for paid Pro. Its launch also mentions access in Gemini for business, while the current Antigravity Gemini Enterprise table lists stand-alone Carbon, Argon, and Hydrogen but marks both Orchestrator variants unavailable. Do not infer Antigravity enterprise access from the launch language. Account access and billed amounts were not tested.
The expensive part of a handoff is often reconstruction
A handoff can add a fresh input and a new prefix, but the larger cost may be reconstructing the reason behind yesterday’s edit. If a worker receives the entire history, it pays for reading material that no longer matters. If it receives only a terse summary, it may miss the negative test that motivated the design. Either extreme can create more retries than the cheaper model saves.
Pass the next worker a bounded bundle: the pinned repository revision, the target artifact, the conditions that must remain true, the evidence behind the plan, which tests have already run, and what would trigger escalation. Include observations, not a claim that every previous conclusion was correct. Make each handoff point to the same operation and resource budget. A second model should not inherit a wider tool scope just because it wrote a more confident plan.
A useful trigger for moving up is evidence of a specific mismatch: an unresolved contract, conflicting test results, or an instruction whose intended effect cannot be determined. Use triggers that can also send work back down once the ambiguity is resolved. This gives routing a measurable purpose beyond choosing the model with the higher score.
Compare four operating modes before adopting a router
Start with tasks whose outcomes are adjudicated independently of the model and keep the repository snapshots, tool set, and time budgets fixed. Split by project family rather than by individual issue when closely related issues share the same code.
| Arm | What it answers |
|---|---|
| Hydrogen throughout | How much can the efficient model deliver by itself? |
| Carbon or Argon throughout | What quality and cost does the stronger baseline achieve? |
| Explicit stage-based handoff | What does an inspectable, persistent routing rule buy? |
| Adaptive routing | Does switching based on observed difficulty beat the simpler rule? |
Treat a product-mode experiment and an API experiment as separate studies. A product-only comparison may answer the purchasing question, while a controlled harness is better for attributing effects. Record which models actually served each stage if the surface exposes that evidence. If it does not, report the product as a whole.
Track tokens, cache usage when visible, turnaround, verification and reviewer minutes, and accepted final changes. Count escalated and abandoned attempts. An adaptive policy can look fast on finished tasks alone simply by dropping difficult cases. Report effectful operations separately from changes applied to an isolated workspace.
The adoption criterion belongs to the task
For a local illustration, a handoff that saves 30 seconds of generation but adds 45 seconds of re-verification makes the task slower. The same handoff may still be useful if it prevents expensive human rework or frees a constrained premium quota. State which outcome matters and test the trade-off on the tasks the team will actually ship.
Hydrogen and Orchestrator are interesting because they let a system change the resolution of reasoning as new evidence arrives. The strongest justification is a better rate of correct, accepted changes within the same budget and authority boundary. A faster model is one way to reach that result; a handoff that preserves the right information is what keeps it possible.