Kanji
・Cloud engineer / freelance ・Born in 1993 ・From Ehime / Living in Shibuya, Tokyo ・5 years AWS experience Profile details
Table of Contents
When Carbon, Argon, Hydrogen, and Orchestrator appear alongside Claude and GPT-OSS in the Antigravity model selector, the first question most engineers ask is “Which one is the strongest?” After running daily implementation and verification loops across repositories, my conclusion is simple: route models by task difficulty, iteration count, and cost of failure .
Carbon, Argon, Hydrogen, and Orchestrator are Antigravity product tier and execution names. Because cross-vendor benchmarks under identical harnesses and full mapping tables to underlying checkpoints are not publicly disclosed for every option, I do not force them into a single composite ranking. Relying on published evaluation conditions and acceptance tests on your own repository is safer than guessing undisclosed numbers.
The sudden expansion of the model selector reflects Antigravity’s evolution from a single-model coding assistant into a multi-model agentic development platform.
Because of this evolution, the dropdown list is not a linear performance ladder. It helps to treat Argon, Carbon, and Hydrogen as functional model roles, and Orchestrator as an execution mode that combines multiple models dynamically . According to the official Orchestrator announcement, on Terminal-Bench 4.1 it maintains task success rates close to standalone Carbon while consuming 24% less quota on average and finishing 19% faster. Still, because evaluation harnesses may differ from your repository and plan limits, validate on real tasks as described below.
Lining up self-reported vendor scores in a single column invites misinterpretation because prompts, tool access, retry counts, and evaluation dates differ. Instead, I evaluate options in three steps:
For coding agents in particular, a single SWE-bench number is never enough—you need to check whether existing tests stay green, scope boundaries are respected, and review follow-up rounds stay low.
The “Best For” and “When to Avoid” columns below represent editorial operational recommendations rather than an official composite benchmark ranking. Public vendor scores cannot be directly merged because test harnesses, tool permissions, and retry budgets vary. Official vendor claims are separated with citations in the next section.
The numbers and performance claims in this section come directly from each vendor’s official publications. They are distinct from the operational routing table above and should not be summed across different evaluation harnesses.
Google DeepMind’s Gemini 4 Carbon Model Card (October 2026) publishes a unified evaluation table covering Gemini 4 Carbon, Gemini 4 Argon, and Claude Opus 5.5 under the same test setup. Below are the key metrics:
These figures reflect Google’s published evaluation conditions, tool configurations, and sampling protocols. Public documentation alone does not confirm whether Gemini 4 Carbon in the model card is identical in system prompt and thinking budget to Gemini 4 Carbon Medium in the Antigravity UI. Treat this table as a model card capability reference , not a guaranteed score inside Antigravity sessions.
Gemini 4 Carbon
Gemini 4 Carbon Medium
Hydrogen and Orchestrator metrics were published separately in the official Antigravity blog announcement rather than the Carbon Model Card. Because the benchmark versions and settings differ, I keep them in a separate table:
Hydrogen’s two scores and Orchestrator’s efficiency gains come from the official Antigravity blog. Orchestrator evaluates a dynamic execution policy that switches between Carbon and Hydrogen as work progresses—so the 24% and 19% figures represent resource and latency savings, not a 24% increase in raw model intelligence.
Google’s official Gemini 3.8 Flash model card also publishes competitive reasoning, knowledge, and coding evaluations, showing why Flash shouldn’t be dismissed as “only for trivial prose.” In real agentic workflows, however, tool-calling discipline, repository rules, and retry caps heavily influence outcomes.
Anthropic’s official release notes state that Claude Opus 5.5 leads in agentic coding, computer use, and knowledge work while cutting tool calls by roughly 40% and halving token usage on terminal tasks compared to its predecessor. Claude Sonnet 5.5 is positioned as over 30% faster than Sonnet 5 and up to 30% more cost-effective across common tasks—supporting the practical split of “Sonnet for daily velocity, Opus for hard problems.”
When tweaking UI copy, regexes, single functions, or initial log triage, response latency dominates total wall-clock time. Start with Gemini 3.8 Flash or Hydrogen, and escalate to Argon immediately if any of the following happen:
Having an explicit escalation trigger keeps latency and quota usage low without getting stuck in retry loops.
This is the standard tier for taking a ticket from requirements to implementation, tests, and review in a single session. Pass requirements as bullet points and explicitly state out-of-scope files, Definition of Done, and test commands to run. If accuracy drops, check for missing context before blaming the model.
Sonnet 5.5 excels when you want fast iterative edits with strong code quality. Between Argon and Sonnet, compare them on similar tickets in your repo and judge by total turns to green CI and diff readability.
For auth, billing, data migrations, production incidents, or wide-scale refactoring, start directly with Carbon or Opus. The key is not blind delegation to a top model, but structured separation of duties:
Adding an independent cross-review step significantly reduces plausible-sounding hallucinations.
Orchestrator dynamically switches bidirectionally between Carbon and Hydrogen within a single execution trajectory . Distinguish this from manually copy-pasting context across separate chat windows and forcing the agent to re-read the repository each time. On Terminal-Bench 4.1, Google reports Carbon-level success rates with 24% lower quota usage and 19% faster completion.
Google does not disclose the exact internal granularity at which conversation history, tool outputs, and file reads are cached or shared during model handoffs. Therefore, do not assume zero context overhead on every switch. For tightly coupled tasks, workflows that repeatedly inspect massive logs, or sessions with frequent human direction changes, sticking to a single model session can be faster and more predictable.
Evaluate GPT-OSS 120B in Antigravity as a managed option inside the Antigravity environment —using it here does not require you to self-host GPUs or maintain isolated inference clusters. Focus your comparison on identical prompts and acceptance tests:
To keep model selection grounded in data rather than gut feeling, log the following across comparable tickets:
Recording these metrics over four weeks quickly reveals repository-specific rules—such as “Flash/Hydrogen handles routine chores,” “Auth changes start on Carbon,” and “Opus runs design verification.”
Model selection in Antigravity is not about chasing the newest name or a single leaderboard number. Separating fast lightweight loops (Flash / Hydrogen), daily implementation drivers (Argon / Sonnet), high-reasoning tiers for critical changes (Carbon / Opus), decomposed workflows (Orchestrator), and open-weight baselines (GPT-OSS 120B) lets you balance engineering quality with quota efficiency.
Public information verified on November 13, 2026. Benchmarks, pricing, and available models in Antigravity may change; always check the latest official model cards and console labels before production deployment.
Master conversation skills with realistic AI phone training and 5-axis feedback. Includes 16-lesson interactive academy and studio AI portraits.
Effortlessly snap receipts for instant AI itemization. Integrated with LINE for seamless group expense tracking and pain-free settlement.
Interactive AI personas (Fortune Teller, Butler, Bartender) host real-time games, challenges, and conversation starters for unforgettable gatherings.
Real-time AI quant analysis of earnings reports, corporate filings, and momentum indicators for global equities and Japanese stocks.
Unbiased healthcare database featuring zero-additive filters and active ingredient cost performance metrics.
Get the complete configuration pack featured in this article: universal AGENTS.md, deterministic hooks.json, Git safety guardrails, and validation scripts tested across Antigravity, Kiro, Claude Code, and Codex.
