Gemini 3.8 Flash Free: Try It on Antigravity or Verdant
Google's Gemini 3.8 Flash tests as a top-tier model. Here's where to try it free, including Antigravity's free tier and the Verdant coding workspace.
What is Gemini 3.8 Flash and why does it matter?
Gemini 3.8 Flash is Google’s latest small, fast model in the Gemini lineup, and it marks a sharp turnaround from its predecessor. Independent benchmark testing on Kingbench 3, an eight-question suite covering front-end coding, 3D work, SVG generation, math, and full agentic tasks, put Gemini 3.8 Flash at 65 out of 80 points (81.25%). That score puts it tied with Qwen 3.8 Max, just under Kimi K5 at 82.5%, and above Claude Opus 4.8 at 80%, landing a “flash” tier model in the top five of the entire leaderboard. For context, the previous version, Gemini 3.5 Flash, scored only 30% on the same benchmark, one of the lowest results tested. That’s a jump of more than 51 points in a single generation, which is unusual for a lightweight model tier that’s normally optimized for speed and cost over raw capability.
TL;DR
- Gemini 3.8 Flash scored 81.25% on an independent eight-question coding and reasoning benchmark, up from just 30% for the prior Gemini 3.5 Flash release.
- The model landed in the top five of the tested leaderboard, ahead of Claude Opus 4.8 and tied with Qwen 3.8 Max.
- Antigravity offers a generous free tier for running Gemini 3.8 Flash, making it a practical option for research and lighter coding work without paying for API access.
- Verdant, an agentic coding workspace available as a desktop app and VS Code extension, now supports Gemini 3.8 Flash and gives new users free trial credits with no card required.
- The model showed a file-overwriting habit, rewriting entire files instead of making targeted edits, a quirk worth watching in real codebases.
- On tasks like game logic and SVG generation, Gemini 3.8 Flash scored notably higher than competing models tested in the same run, including a perfect score on a bow-and-arrow game simulation.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
How does Gemini 3.8 Flash perform on real coding tasks?
The benchmark used to evaluate the model, Kingbench 3, runs eight distinct tests: an elevator simulation with queuing logic, a 3D contact lens case built in Three.js, a folding table animation, an SVG drawing of a panda eating a burger, a bow-and-arrow game with a leaderboard, a permutation math problem, a full agentic pipeline (generating a dataset, fine-tuning a small model, and serving results through a local web UI), and a 3D wristwatch simulation with dual time zones.
Gemini 3.8 Flash’s standout results came on the 3D contact lens case (a 9 out of 10), the SVG panda (an 8), and the bow-and-arrow game, where it scored a full 10, correctly handling all four targets, responsive aiming, and accurate leaderboard tracking. It also nailed the hard math question, landing on the exact correct answer of 2460, and completed the full agentic test end to end, generating a dataset, fine-tuning a Gemma 2B model locally, and serving results through a working web UI.
It wasn’t flawless. The folding table test, which involves smooth 3D geometry animation, tripped it up (a 5 out of 10), and the wristwatch test, one of the hardest in the suite, came in at 6 out of 10 due to rough handling of the second time zone. Still, for a model in the “flash” tier, typically the cheapest and fastest option in a lineup rather than the most capable, these numbers are well above what’s expected.
What are the downsides of Gemini 3.8 Flash?
The most notable issue found in testing was a tendency to overwrite entire files rather than making targeted, minimal edits. Ask the model for a small change and it may rewrite the whole file instead of patching just the relevant lines. In a real codebase, that means every diff needs a careful review to make sure nothing unrelated got quietly changed or removed. This is a habit that wasn’t as pronounced in earlier Gemini models, so it appears to be new to this release.
This isn’t a dealbreaker, but it does mean workflows built around Gemini 3.8 Flash benefit from small, frequent commits and actual diff review rather than blind trust in agentic edits. Pairing the model with a workspace that includes planning and verification steps helps offset this risk.
Where can you access Gemini 3.8 Flash for free?
Two main routes stand out for trying Gemini 3.8 Flash without paying full API rates.
Antigravity is the more straightforward option. Flash-tier models run quickly there, and the platform’s free tier is described as generous enough for regular use, particularly for research work and lighter coding tasks. Because Gemini 3.8 Flash already tests as a strong, fast model, running it through a free tier makes it a genuinely useful option for students or anyone who wants capable AI assistance without a subscription commitment.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
Verdant is the second option, and it takes a different approach. It’s an agentic coding workspace available both as a desktop application and as a VS Code extension. Its core feature is orchestration: running multiple AI agents in parallel, each isolated in its own workspace so they don’t interfere with one another. Verdant also includes a plan mode, where the agent lays out its intended steps before making changes, along with separate verification and review agents that check completed work afterward.
Gemini 3.8 Flash is now available inside Verdant, and new users get free trial credits without needing to enter payment information first. Given the file-overwriting behavior noted above, Verdant’s built-in plan and review steps are particularly useful here, since they add a checkpoint before and after the model touches any code.
Is Gemini 3.8 Flash worth using over other models?
For a model in the flash tier, the benchmark results suggest yes, particularly for cost-sensitive or speed-sensitive use cases. Landing in the top five of a leaderboard that includes larger, more expensive models like Claude Opus 4.8 is a significant result for a model designed to be fast and cheap rather than maximally capable. The jump from 30% to 81.25% between generations also suggests Google made substantial improvements to reasoning and code generation in this release rather than just tuning for speed.
That said, the file-overwrite quirk is a real practical concern for anyone using it in production codebases. It’s not a reason to avoid the model, but it is a reason to build in review steps, whether through manual diff checking or a workspace like Verdant that includes verification agents by default.
Frequently Asked Questions
What is the difference between Gemini 3.8 Flash and Gemini 3.5 Flash?
Gemini 3.5 Flash scored 30% on an independent coding and reasoning benchmark, one of the lowest results recorded for any tested model. Gemini 3.8 Flash scored 81.25% on the same benchmark, a jump of more than 51 percentage points, placing it in the top five of the leaderboard rather than near the bottom.
Can I use Gemini 3.8 Flash without an API key or paid subscription?
Yes. Antigravity offers a free tier that supports Gemini 3.8 Flash, and Verdant gives new users free trial credits without requiring a card upfront, making both viable ways to try the model at no cost.
What is Verdant and how is it different from a normal code editor?
Verdant is an agentic coding workspace available as a desktop app and a VS Code extension. Its main feature is running multiple AI agents in parallel, each in an isolated workspace, along with a plan mode that outlines intended changes before execution and review agents that verify completed work.
Does Gemini 3.8 Flash have any known issues?
The main issue found in testing is that it tends to overwrite entire files instead of making small, targeted edits, even for minor requested changes. This makes careful diff review important when using it in real projects.
How does Gemini 3.8 Flash compare to Meta’s Muse Spark 1.3?
In the same benchmark run, Gemini 3.8 Flash scored 81.25%, while Muse Spark 1.3 scored 71.25%, actually a regression from Muse Spark 1.2’s 76.25%. Gemini 3.8 Flash outperformed Muse Spark 1.3 on tests like the SVG panda drawing and the bow-and-arrow game, though Muse Spark 1.3 set a new high score on the 3D wristwatch test.
