Published 2026-02-14.
Time to read: 4 minutes.
llm collection.
In 2005, Google launched Google Code Assistant (GCA), which could be used as an plugin for IDEs that added agentic programming capability, and could also be used from a command line.
On November 18, 2025, Google launched a public preview of Antigravity, their agent-first development platform. Unlike AI Studio, Antigravity is designed to bridge the gap between their high-reasoning models and your local environment.
Antigravity uses the Model Context Protocol (MCP) to communicate with a local
server on your machine. Consequently, it can work with your local file
structure, request to edit files, and (with your permission) trigger local shell
commands like go test ./....
Pricing includes a more generous free tier than the Gemini free Pro tier.
Antigravity is a preview, not a released product
Installation
I downloaded the Windows version for x64.
One of the installation steps requires the user to indicate the level of granularity for permissions. I accepted the default, review-driven development.
I disallowed Google to use my interactions for any purpose.
Suitability
antigravity shows a pro (high) and pro (low) model
In Google Antigravity, the “High” vs. “Low” settings for Gemini 3 Pro act as
a toggle for Reasoning Depth (Thinking Level). Since you’re working on the
sc_router gateway, this choice is critical for managing your quota and getting
valid code.
Gemini 3 Pro (High) vs. (Low)
| Feature | Pro (High) | Pro (Low) |
|---|---|---|
| Primary Use | Complex Refactoring, Architectural Logic, Deep Debugging. | Standard Coding, UI tweaks, unit test boilerplate. |
| Quota Cost | Expensive. Burns through your “fuel tank” quickly (sometimes 50–100x a standard prompt). | Efficient. Lower cost per prompt; keeps you under the limit longer. |
| Logic/Thinking | Uses a “Deep Think” or “Chain of Thought” process before outputting. | Direct response; faster but more prone to missing edge cases. |
| Context | Full access to the 1M+ token window. | Full access, but handles less “active” detail at once. |
How to use them
Because you mentioned that Opus and Minimax failed due to complexity, Pro (Low) will likely fail you too. It will behave more like a “standard” LLM and struggle with the logic.
My recommendation for your current session:
- Use Pro (High) for the core refactoring steps where logic is dense.
- Toggle to Pro (Low) or Flash for the “busy work”—like writing the docstrings, simple getters/setters, or repetitive test cases.
- Watch the “Fuel Tank”: In Antigravity, usage isn’t per-message; it’s per
“unit of work.” A single
refactor this modulecommand in High mode can exhaust a free-tier weekly quota in just a few turns.
Subscription Note
If you are seeing these options, you are likely in the Public Preview. If you upgrade to Google AI Pro/Ultra, your Antigravity limits typically refresh every 5 hours instead of the weekly reset faced by free users.
i do not recall gca offering two pro settings. is the gca pro model equivalent to high?
In 2026, the landscape has shifted slightly with the introduction of Gemini 3, which is why you’re seeing these specific “High/Low” toggles in the newer Antigravity IDE that don’t appear in the more traditional Gemini Code Assist (GCA).
Is GCA Pro equivalent to “High”?
Not exactly. Here is the distinction:
-
Gemini Code Assist (GCA): This uses Gemini 2.5 Pro or 3 Pro in a “Balanced” or “Dynamic” state. It doesn’t give you a manual toggle because it is designed to manage its own reasoning tokens to stay within the 240 chat requests/day limit. It is effectively “Medium-High”—it will “think” if the code is hard, but it won’t go into a deep, multi-minute reasoning loop unless it’s an Enterprise version.
-
Antigravity “Pro (High)”: This explicitly triggers the “Deep Think” mode of Gemini 3 Pro. This mode is a “Chain of Thought” specialist. It spends significant compute time (and quota) simulating a developer’s thought process before it writes a single character. This is what you need for the
sc_routercomplexity that failed other models. -
Antigravity “Pro (Low)”: This is a lower-latency version of Pro. It has the same general knowledge as the High version but skips the “internal monologue” to save time and credits.
Why you see two settings in Antigravity but not GCA
Antigravity is Google’s newer “agent-first” platform (similar to Cursor). Because it is built for agents that work autonomously, Google gives the user a “gas pedal” (High/Low) to decide how much compute to spend on a task.
GCA is an IDE extension for VS Code/JetBrains. It is built for a more traditional “autocomplete and chat” workflow where Google manages the model backend to ensure you get a consistent response speed.
Comparison for Your sc_router Task
| Setting | Reasoning Style | Best For | Quota Impact |
|---|---|---|---|
| GCA (Standard) | Reactive | General debugging & small refactors. | Fixed daily requests (240). |
| Antigravity (Low) | Direct | Boilerplate, unit tests, simple logic. | Moderate credit burn. |
| Antigravity (High) | Chain-of-Thought | Complex architectural refactors. | Heavy credit burn. |
Recommendation
If you decide to move your refactoring to Antigravity once your free access resets:
-
Use Pro (High) for the specific parts where you need to coordinate the high-performance Git LFS logic with the Go router gateway.
-
Switch to Pro (Low) or Flash for writing the documentation and the boring parts of the implementation.
Since you’ve seen the “Usage limit reached” message, it means you’ve hit the cap on the “Pro” engine itself. When it resets, you’ll have a choice: stay in the chat (GCA-style) or move the codebase into Antigravity to use that “High” reasoning toggle.
It seems like Antigravity either offers too much, at a very high price, or not enough. In contrast CGA is working very well.
I completely understand that sentiment. Antigravity can feel like trying to fly a jet when all you wanted was a reliable car—it’s heavy on “agent orchestration” and “missions,” which adds significant cognitive overhead and cost if you just want to write code.
Since Gemini Code Assist (GCA) has the right “vibe” and reasoning level for
your sc_router project, here is how the pricing and quota reality looks for
you in 2026:
GCA Pricing & “Pro” Access
The “just right” feeling you had with GCA Pro is a result of its Balanced Reasoning and 1-million-token context window. To keep that consistently without hitting the free-tier “Usage limit reached” message, you have two real paths:
| Plan | Cost | Quota (Pro Models) | Best for sc_router
|
|---|---|---|---|
| Free Tier | $0 | ~240 chat / 6,000 code requests per day. | Good for small fixes; bad for deep refactors. |
| Google AI Pro | $19.99/mo | Higher/Uncapped daily requests for GCA. | The Sweet Spot. Keeps the GCA interface you like. |
| Pay-As-You-Go | Per Token | Unlimited (based on credit). | Only if you use the Gemini CLI or API. |
Why Antigravity feels like “Too Much”
Antigravity is Google’s attempt to compete with “agentic” IDEs like Cursor. It introduces:
-
High/Low Toggles: These are compute-intensity switches. “High” is basically “Deep Research” for code, which is why it’s so expensive and often “overthinks” simple Go logic.
-
Agentic Overhead: It wants to “plan” and “verify” rather than just giving you the code you asked for.
Recommendation
If GCA Pro was working perfectly for your Golang gateway, don’t move to Antigravity. The stability issues and “High/Low” complexity will likely frustrate you.
