When was Antigravity 3.7 launched, and what changes were made?

Google Antigravity 3.7: I Spent 24 Hours With It and Honestly, I'm Still Processing What I Saw

Okay so tech moves fast, we all know that. But what happened this week broke my brain a little.

I was literally three weeks into getting comfortable with Antigravity 3.6 Flash, still automating a few of my projects with it, when out of nowhere Google dropped Antigravity 3.7 (Gemini 3.7 Flash) yesterday, August 13, 2026. No warning. Just... there.

I've been running Antigravity in my daily workflow for a while now — coding, deploying AI agents, the usual grind. So the second this update hit, I didn't wait around. I dove straight in. For the last 24 hours I've been hammering it with everything I've got: heavy reasoning tasks, UI-to-code conversions, the works.

If you're on the fence about whether this is a genuine leap forward or just another minor patch dressed up as a big release — trust me, my hands-on testing is about to answer that for you.

The Launch Speed Alone Is Kind of Unhinged

Let's start here because it genuinely threw me off. Big LLM releases usually take months to cook. Antigravity 3.6 Flash to 3.7 in just three weeks? That's not a normal release cadence — that's Google going full send on owning the agentic AI market before anyone else gets a foothold. The moment I got access last night, I didn't even finish my coffee before I was elbow-deep in the new features.

1. My Favorite Feature: Live-Testing "Thinking Levels"

The single biggest, most genuinely useful change I found in 3.7 is control over Thinking Levels. Until now, you'd throw a prompt at the model and it would think however much it felt like thinking. Now there's an actual UI toggle: Low, Medium, High.

I ran three completely different scenarios to see what this actually does:

Low Thinking: I gave it a basic real-time chat task and some data formatting. Response time was instant — like, blink-and-you-missed-it fast. Perfect for stuff where you need an answer now and there's no heavy backend logic involved.

Medium Thinking (the default): This is where I ran my everyday coding tasks and API integrations. Honestly the sweet spot between speed and actually understanding what I asked for.

High Thinking — this is the real game changer: I threw a genuinely nasty math algorithm at it plus a broken codebase that needed fixing. It didn't answer right away. It sat there for a few seconds (actual deep thinking), ran tools in the backend, caught and fixed its own mistakes mid-process, and the final code it handed me ran clean on the first try. No bugs. First shot.

2. Coding and Agentic Tasks: Is the Improvement Real?

As a developer, the upgrade I felt most in my bones was the agentic capability. I handed the Antigravity Agent an old web scraping project — one where a bunch of the target sites had changed their structure since I last touched it — and the planning it showed was honestly impressive.

• Self-correction: When a selector failed mid-run, it didn't just stall out. It read the error, adapted, and found another way through.
• First-pass accuracy: I barely had to hit "regenerate." Whatever came out on the first attempt was already about 90% production-ready.

And the numbers back this up too — Google's claiming a jump from 49.0% to 65.3% on DeepSWE v1.1, the software engineering benchmark. And yeah, in my own testing, that gap was noticeable.

3. Design-to-Code: From a Figma Screenshot to a Working App

I put this thing through one more stress test. Took a screenshot of an old Figma design and told Antigravity 3.7: "Give me a 1:1 matching React build of this."

Older models used to mangle the layout or completely botch the padding and margins. Not this one. The CSS and Tailwind code it generated was scarily accurate — this is what people in the industry call "design parity," and 3.7 clearly has it. It's scoring 1588 Elo on the WebDev Arena benchmark now, which tells you this thing is about to save frontend developers a serious chunk of their week.

4. Your Wallet Will Thank You: A 50% Price Cut

As a developer and someone building things on the side, cost is always a factor I can't ignore. This is where Google really won me over. Alongside the 3.7 release, they rolled out promotional pricing running through December 31, 2026.

Prices are literally cut in half:

• Input tokens: $0.75 per 1 million tokens
• Output tokens: $3.75 per 1 million tokens

Last night I was processing a genuinely heavy volume of data and my bill came in way under what I expected. If you're building AI apps at scale, this is a window you don't want to miss — cheap testing and cheap scaling for the next few months.

5. Under the Hood: 1 Million Context and a Massive Output Limit

I wanted to stress-test the context window, so I uploaded an entire code repository — roughly 300,000 tokens — in one go. With a knowledge cutoff of March 2026, it still understood my whole project without missing a beat.

Here's the part that really got me though: the output limit now goes up to 64,000 tokens. That's not just "write me a snippet" territory anymore — you can get it to write an entire long script or a full-length report in one uninterrupted pass.

Final Take: Should You Actually Use This?

After 24 hours with it, here's where I land: Google Antigravity 3.7 is not a minor version bump. This is a genuine leap forward for AI agents. The Thinking Levels control alone is one of the most useful, most thoughtfully designed features I've seen in a release like this.

If you're a developer, a data analyst, or just someone who likes messing around with AI tools — try this one immediately. It's fast, it's cheap right now, and it's noticeably smarter than what came before it.

Lovable.dev Review 2026: My Honest Take After Actually Using It

Grok Build Was Secretly Uploading Your Entire GitHub Repo-Here's the Full Story