Google Antigravity adapts the model during tasks
Orchestrator in Google Antigravity switches between Gemini 4 Carbon and Hydrogen throughout a session. Google reports gains in speed and quota, with results close to Carbon on Terminal-Bench 4.1.
On November 12, 2026, Google introduced Orchestrator, a dynamic model selection option in Antigravity reserved for Google AI Ultra subscribers, while the Hydrogen model is deployed separately for Pro and Ultra subscribers. The idea is simple to explain, but harder to maintain in production: a single agent can switch from Gemini 4 Carbon to Gemini 4 Hydrogen during a session, depending on planning, execution, or repair needs. On Terminal-Bench 4.1, Google reports results close to Carbon, with 24% less quota and 19% faster execution time. At the same time, Innodisk expands its AI data center component offering with SSDs, DDR5, and CXL, proving that the AI battle is also fought in the less visible hardware layers.
When the model changes mid-sentence
Orchestrator does not just choose a model at the start and stick to it at all costs. Google explains that the tool can switch from one model to another during the same task, allowing it to adjust the reasoning level depending on the moment: initial investigation, code writing, testing, then resuming in case of blockage. Carbon is used for planning or recovery phases, while Hydrogen takes over for fast execution. This mechanism is more flexible than a static router, which decides once and for all at launch.
This detail matters for developers working on long and changing tasks, especially debugging and feature deployment. Google targets a well-known weak point of agents: when the real difficulty only appears along the way, a fixed allocation can waste time and burn quota unnecessarily. The system thus resembles less a binary switch and more a conductor who decides, measure by measure, who plays the next part. The kind of flexibility that makes product teams love diagrams, then haunt them during integration.
Hydrogen, Carbon and better-shared use cases
Hydrogen is presented as the fast execution model, while Carbon retains the role of deeper reasoning and correction. Google also specifies that the Antigravity SDK allows developers to explicitly assign each model to a given step, instead of leaving Orchestrator to arbitrate alone. In an open-source ticket triage demonstration, Hydrogen handles the incoming flow, searches the code, asks clarification questions, and iterates on the interface, while Carbon takes on heavier refactors, new features, and security checks.
This separation follows a logic that has become familiar in agentic AI tools: not all work deserves the same computing power. A fast model is often enough to move forward, but a more capable model remains useful when it is necessary to pick up the thread, correct a hypothesis, or secure a decision. Google adds that other models and subscription plans will be supported later for Orchestrator, without giving a precise timeline. At this stage, the interest is not only technical. It is also economic: less quota consumed means less friction for teams multiplying trials.
What Google opposes to Swarm
Orchestrator arrives after the preview of Gemini 4 Carbon and /swarm, released on November 10, 2026. But the two tools do not solve the same problem. /swarm distributes a task among several agents on distinct files, packages, or questions, with verifiers responsible for reviewing diffs and running tests. Orchestrator, on the other hand, works within a single evolving session, switching models when the context requires it. The first organizes work between agents; the second arbitrates between models.
This distinction may seem subtle, but it outlines an important evolution of AI workflows. Model selection no longer happens only at the initial prompt. It becomes an architectural variable, on par with task breakdown or error feedback management. For companies integrating these systems into their internal tools, the potential gain is clear: better adaptation to real tasks, without imposing the cost of the heaviest model at every step.
Innodisk also strengthens the AI hardware layer
On the same November 12, 2026, Innodisk announced the expansion of its AI data center portfolio, with enterprise SSDs, DDR5 memory, and CXL modules aimed at cloud providers and neocloud operators. The company highlights the 3TS9-P, 4TS2-P, and 5TS-P series, available for sampling and customization, as well as a memory range designed for long-context inference workloads and HPC calculations. The promise is clear: more bandwidth, more capacity, and less integration hassle.
Regarding storage, Innodisk announces up to 14 GB/s sequential read and 10 GB/s sequential write for the 5TS-P series, with capacities from 1.92 TB to 7.68 TB for the E1.S model. The company also mentions DDR5 memory at 12,800 MT/s, offered from 32 GB to 128 GB, as well as a CXL module of 64 GB, 96 GB, or 128 GB capable of adding up to 32 GB/s bandwidth. These figures matter because they show where AI infrastructure competitiveness is now played: not only in accelerators but also in storage, memory, and thermal management.
What this says about the AI market
Google and Innodisk do not sell the same product, but they respond to the same reality: AI systems are becoming more modular, more costly to orchestrate, and more sensitive to bottlenecks. On one side, software tries to better distribute effort between reasoning and execution. On the other, hardware tries to absorb ever larger volumes of data, memory, and traffic. In both cases, the goal is to reduce friction, because AI dislikes downtime and surprises. Well, almost never.
In the short term, the most concrete signal remains real adoption: at Google, the question will be whether Orchestrator delivers on its promises beyond demonstration; at Innodisk, whether the SSDs and memory modules sampled pass cloud customer qualifications. These validations, more than announcements, will say if these building blocks find their place in production infrastructures.
Key points
- Orchestrator arrives on November 12, 2026, in Antigravity.
- Google announces 24% less quota on Terminal-Bench 4.1.
- Hydrogen and Carbon are available via Google AI Pro and Ultra.
- Innodisk presents its 3TS9-P, 4TS2-P, and 5TS-P SSDs.
- CXL adds up to 32 GB/s bandwidth at Innodisk.
By the numbers
- 24% — less quota, according to Google, on Terminal-Bench 4.1.
- 19% — completion time gained, according to Google, in one session.
- 14 GB/s — maximum sequential read announced by Innodisk for the 5TS-P series.
- 12,800 MT/s — throughput of Innodisk’s DDR5 12800 Multiplexed Rank DIMM.
- 32 GB/s — additional bandwidth announced for Innodisk’s CXL module.