Course contents
Choosing a Model: Carbon, Argon, Flash, and Claude
Pick the model by the job: Gemini 4 Carbon for coordination and verification, Gemini 4 Argon as the everyday default, Gemini 3.8 Flash for routine edits, and Claude when you want a second model family. Then see what each costs per token on the API.
Antigravity's model picker is a budget decision as much as a quality one. Stronger models reason more, use more tokens and draw down your quota faster. The skill is matching the model to the cognitive demand of each task, and of each role inside a task, instead of picking one favorite and using it everywhere.
What's in the picker, as of November 2026
| Model | Plans | Effort (default) | What Google says it's for |
|---|---|---|---|
| Gemini 4 Carbon | Google AI Ultra, enterprise | Low, Medium, High (High) | Exploring thoroughly and running extended verification loops: coordinator, architecture, security-auditing and verification roles |
| Gemini 4 Argon | Paid Google AI Pro, Ultra, enterprise | Low, Medium, High (High) | "An ideal choice for most advanced engineering and knowledge work tasks" |
| Gemini 3.8 Flash | Every plan, including Free | Low, Medium, High, plus Fast (Medium) | The most intelligent Flash model, for long-horizon software engineering and agents |
| Claude Opus 5.5 and Claude Sonnet 5.5 (Thinking) | Paid Google AI Pro (not trials), Ultra, enterprise | Not listed | A second model family for panels and reviews |
Older models are on their way out. Claude Opus 4.6, Claude Sonnet 4.6 and GPT-OSS 120B are scheduled for removal on November 18, 2026, and Gemini 3.7 Flash, Gemini 3.6 Flash and Gemini 3.1 Pro remain only for enterprise accounts. Usage is metered in two pools, one for Gemini models and one for Claude and GPT models, each with a weekly limit and a five-hour limit.
On a Google AI Ultra plan, which Gemini model fits each job?
Debug a race condition you can reproduce
Effort is a second dial
Most models offer Low, Medium and High effort, and Flash adds a Fast option. Higher effort trades response speed for deeper reasoning: use it for architecture and debugging, and drop it for quick edits. Your model choice persists across a conversation, and switching mid-run takes effect after the current turn finishes or is cancelled. In the CLI, /model and /effort change them, and agy models lists what your account can use.
Plan limits: what your plan buys
| Plan | Antigravity usage | Models it adds |
|---|---|---|
| Free and Google AI Plus | "A meaningful quota, refreshed weekly" | Gemini 3.8 Flash |
| Google AI Pro | "A high, generous quota, refreshed every five hours until the weekly limit is reached" | Gemini 4 Argon; Claude Opus 5.5 and Sonnet 5.5 on non-trial plans |
| Google AI Ultra, $100 tier | 5x Pro's token rate limit | Gemini 4 Carbon |
| Google AI Ultra, $200 tier | 20x Pro's token rate limit | Gemini 4 Carbon |
| Gemini Enterprise (business) | Under Google Cloud terms | Carbon through Gemini for business, plus the older enterprise-only models |
Limits track agent work, not prompt count: in Google's words, "more complex tasks consume more quota." At I/O in May 2026 Google priced Google AI Pro at $20 a month and the two Ultra tiers at $100 and $200 (the top tier down from $250); check current prices for your region. When you hit a limit, the AI Credit Overages setting decides what happens: wait for the quota to refresh, or spend purchased AI credits. Bringing your own API key for extra App limits isn't supported.
You're on a non-trial Google AI Pro plan with AI Credit Overages set to never. Mid-morning you hit the five-hour limit on Gemini models. What happens?
Paying per token: the SDK and the API agent
Plan quotas cover the App, the CLI and the IDE. Agents you build with the Antigravity SDK, and the managed Antigravity agent on the Gemini API (antigravity-preview-09-2026, built on Gemini 3.8 Flash), bill per token at Gemini API rates instead. As of November 2026 those are launch prices, and two changes are scheduled: Gemini 4 Carbon and Gemini 4 Argon double on December 1, 2026, and Gemini 3.8 Flash doubles on January 1, 2027. Gemini 3.5 Flash-Lite, the cheapest current model, is API-only and not in Antigravity's picker. API access to Carbon is being phased in, so check AI Studio before you plan around it.
Carbon and Argon share one per-token price. Gemini 3.8 Flash also doubles, to $1.50 in and $7.50 out per million, on January 1, 2027. One request here is one model call; a single agent task makes many.
Move the controls and watch what changes.
With the starting workload (50,000 input and 4,000 output tokens per call), which part of the Gemini 4 bill is bigger?
Carbon and Argon show the same monthly cost in the widget. Why might Carbon still cost more per finished task?
Plain-English definitions from the AI Coding Dictionary.