Google Rolls Out Orchestrator in Antigravity, Pairing a Carbon or Argon Planner With Hydrogen; Two Google Posts Name Different Terminal-Bench Versions

Overview

Google is rolling out Orchestrator, a new model-selector option in its Antigravity coding tool, according to the Antigravity blog, which dates the post November 12, 2026. The blog says Orchestrator “pairs Hydrogen with our strongest models for long-horizon tasks” and that it is being rolled out alongside Gemini 4 Hydrogen. Google’s documentation describes two options, Orchestrator (Carbon) and Orchestrator (Argon), named for the planner model each uses. Google says Orchestrator reached “near-Carbon task success rates” while using 24% less quota and finishing 19% faster. Google’s own pages disagree on which Terminal-Bench version produced that result, on which planner options exist and on whether enterprise customers get the feature. WinBuzzer, SiliconANGLE, The Decoder and 9to5Google also covered it.

How Orchestrator works, according to Google’s documentation

The Antigravity models page says Orchestrator “pairs two Gemini 4 models in a single conversation”: a larger planner model does the planning and review, and Gemini 4 Hydrogen does the implementation. The page says each message starts with a read-only planning phase, after which Hydrogen applies file edits, writes code, runs builds and tests, and checks the results against the plan. It lists conditions that return control to the planner, among them “a verification step that fails twice”, and says Hydrogen reports progress to the planner on longer tasks. The page says these hand-offs happen “within a single response”.

Google contrasts this with a router. The Antigravity blog describes “a static “model router”” that guesses task difficulty from the first prompt, and the models page says Orchestrator “isn’t a router”. Orchestrator “runs with a fixed configuration”, so no thinking level can be chosen for it; selecting Carbon, Argon or Hydrogen directly restores that control. The page says to select it with /model or with agy --model "Orchestrator (Carbon)", calls it the “recommended default” for most day-to-day development, and recommends Carbon or Argon directly for work where “every intermediate step needs maximum reasoning depth”. It adds that Orchestrator keeps a single agent in the conversation and points to subagents, Swarm, Boost or Teamwork for parallel work. The blog adds that the Antigravity SDK can wire Hydrogen and Carbon together explicitly when developers want programmatic control over which model handles each role.

What Google reports

The Antigravity blog says: “On Terminal-Bench 4.1, Orchestrator reached near-Carbon task success rates while, on average, consuming 24% less quota and completing 19% faster.” The Google blog post on Hydrogen carries the same figures but names Terminal-Bench 4.0. Neither page explains the difference. The benchmark’s v4.0.0 and v4.1.0 releases on GitHub carry timestamps of 2026-08-26 and 2026-11-03 (UTC). The v4.1.0 notes list “Tasks modified (66)” and “Tasks added or removed: None”. The Herald does not know which version Google used. SiliconANGLE and The Decoder both say 4.0, as the Google blog does; WinBuzzer notes that the two Google posts differ.

None of the pages cited here gives an absolute success rate, a number of runs or a quota unit for Orchestrator. Neither Google post says whether the figures cover the Argon-planned option. The Antigravity blog does not name the baseline for the 24% and 19% figures; WinBuzzer writes that they are measured against “its Carbon planning model”. The Decoder renders the 24% as “fewer resources”, while Google says quota. WinBuzzer adds that Antigravity meters plan quota by the work performed, so the percentage “supplies no fixed dollar saving”. 9to5Google quotes a description of Orchestrator as “Carbon that may offload routine steps to Hydrogen. Same Carbon-level quality, fewer tokens used”, whereas the Google posts say “near-Carbon task success rates”. None of the cited sources reports a measurement of Orchestrator made outside Google; SiliconANGLE, The Decoder and 9to5Google restate Google’s figures, and WinBuzzer labels them as Google’s.

Where the pages differ on availability

• Planner options. The Antigravity blog and the models page list Orchestrator (Carbon) and Orchestrator (Argon). The Antigravity changelog entry dated November 12 describes Orchestrator “with Gemini 4 Carbon”. WinBuzzer, SiliconANGLE and The Decoder describe it as using Carbon and Hydrogen; WinBuzzer’s article text does not mention Argon.
• Plans. The blog says the Carbon option is for Google AI Ultra subscribers and the Argon option for Ultra and for paid Google AI Pro. The models page agrees and adds Pro trials on eligible premium devices. WinBuzzer’s summary says Orchestrator is “reserved for Ultra in the model documentation”, which matches the Carbon option only.
• Enterprise. The blog says “Enterprises can access both in Gemini for business”. The Antigravity enterprise page lists standalone Carbon, Argon and Hydrogen as supported and marks both Orchestrator options as not supported. WinBuzzer reports the same mismatch.
• Standalone Hydrogen. The blog and the outlets above say Hydrogen is for paid Pro and Ultra subscribers in Antigravity; the models page lists it for Free and Google AI Plus accounts too, which WinBuzzer also notes. The changelog says new versions “roll out gradually and may take a few days to reach all users”.

What is not known

The cited sources do not say which Terminal-Bench version, task set or number of runs underlies Google’s figures, what baseline Google used, or how much quota a typical task consumes. They do not say whether the Argon-planned option was measured, or how the launch post’s enterprise statement relates to the enterprise page.