Google has released its internal Antigravity coding agent from isolation and integrated it into the Gemini API as a managed agent. This means developers can now call the agent not just through a standalone IDE, but directly via the Gemini API and Google AI Studio—free during the preview phase, with a distinctive billing model.
At a glance
- Antigravity moves into the Gemini API: The agent now runs as a managed service via the Interactions API, available for both free and paid projects
- Sandbox compute is free (preview): CPU, memory, and Linux sandbox execution are not charged during the preview—only model tokens are billed
- Runs on Gemini 3.8 Flash by default: Google uses its latest Flash model as the standard, with older versions configurable
- Token costs can be substantial: Complex workflows can consume 3–5 million tokens per interaction, costing up to around $5 depending on the task
What Antigravity does in the sandbox
The agent follows a clear pattern: it plans a task, executes code, observes results, and iterates until the task completes or a limit is hit. Inside the sandbox, it has access to code execution, Google Search, URL fetching, and file management. This sounds like a normal chat request—but it's fundamentally different. While a simple prompt delivers one answer, an Antigravity interaction can loop through the agent cycle multiple times, burning tokens as it writes a script, runs tests, reads output, modifies the file, and tries again.
Google's documentation warns explicitly: complex workflows can reach 3 to 5 million tokens in a single interaction, with estimated costs of up to about $5 per task.
Pricing structure: free preview, then higher rates
| Model | Input (through 12/31/2026) | Output (through 12/31/2026) | Input (from 1/1/2027) | Output (from 1/1/2027) |
|---|---|---|---|---|
| Gemini 3.8/3.7/3.6 Flash | $0.75/M | $3.75/M | $1.50/M | $7.50/M |
Here's the hook: sandbox compute itself is free—at least during the preview. That significantly lowers the barrier to entry. If you're testing a workflow that installs packages, runs a build, fixes a failing test, and generates a report, you don't have to worry about the Linux machine's meter ticking. Only token usage stays in focus—and as noted, that can get expensive fast.
Context: Gemini 3.8 Flash and Google's speed strategy
Antigravity runs by default on Gemini 3.8 Flash, which Google released on September 2, 2026—the third Flash model in six weeks. The signal is unmistakable: Google is betting on fast, cheap models rather than just the largest ones. According to Google, Gemini 3.8 Flash beats both GPT-5.6 Terra and Claude Sonnet 5 on the Terminal-Bench 2.1 coding benchmark.
What this means for German developers and enterprises
For German tech teams, this is a low-friction opportunity to experiment with a productive AI agent without immediate high infrastructure costs. Free sandbox compute during the preview reduces financial risk when prototyping. However, token budgets deserve serious attention—anyone planning complex automations should calculate that a single agent interaction can easily land in the five-digit token range. For organizations already invested in Gemini, the API integration is a natural next step; for others, it's a test bed to see whether agent-based workflows fit their processes.
Sources
Editorially owned by Ideal Syka. Sources and method: Newsroom & method. Tips and corrections: ai@i6eal.de.
