A few hours ago, Google released Gemini 3.7 Flash, calling it their “most intelligent workhorse model yet for coding and agents.”
We’ve been testing it since the moment it dropped, and as of right now it’s the model powering the Nowa agent. That means two things for you, starting today:
- The agent got noticeably smarter, especially at building UIs, debugging, and finishing multi-step tasks.
- Your daily limits just doubled on every plan, at the same price. No action needed on your side.
Let’s go through both.
What “smarter” actually means
Benchmarks can get technical, so here’s the short version of what stood out to us, and why it matters when you’re building in Nowa.
It’s the top-ranked model at turning descriptions into working screens. On Code Arena, a public leaderboard where AI models go head-to-head on building real, functional interfaces, 3.7 Flash currently sits at #1 ahead of every other model. In practice, that means screens that look right on the first try, and layouts that actually match what you described.
It’s much better at following a design. Give it a screenshot, an image, or a design system, and it sticks to it far more faithfully than the previous model. If you’ve ever pasted a reference and gotten back something “close but not quite,” this update is for you.
It doesn’t lose the thread on long tasks. On DeepSWE, a benchmark for long, multi-step engineering work, it jumped from 49% to 65% compared to 3.6 Flash. Fewer moments where the agent forgets what you asked three steps ago, and fewer retries to get there.
It writes more correct code the first time. Production code quality went from 34.4% to 43.6% on FrontierCode, which for you simply means less debugging and faster results.
65% 49%
43.6% 34.4%
All figures compare Gemini 3.7 Flash with 3.6 Flash. Source: Google’s announcement.
The part we’re most excited about: your limits doubled
Google launched 3.7 Flash at half the cost of the previous model. We could have quietly kept that saving. Instead, we’re passing all of it to you.
Effective immediately, at the same price. Nothing to claim, nothing to upgrade — open Nowa and it’s already there.
And honestly, in day-to-day use it can feel like more than double. The new model is also more token-efficient, meaning it gets the same task done using fewer tokens. Cheaper tokens, and fewer of them per task. It compounds. The same goes for any extra credits you buy. They’ll now stretch across roughly twice the work they did yesterday.
One more thing
Alongside the model swap, we shipped improvements to the agent itself, with a particular focus on UI generation. Between our updates and the new model, this is the biggest single-day jump in quality the Nowa agent has ever had.
This is one of our favorite things about building Nowa: when the frontier moves, you feel it the same day, not months later.
Go build something. We can’t wait to see it.
Start building →
Happy building,
The Nowa team
