The Vibe Father
Gemini 3.6 Flash is almost certainly real and in internal testing — a model ID reading "gemini-3.6-flash-tiered" surfaced inside Google's Antigravity dev tool early this morning, July 21, 2026, spotted by tester @Lentils80 at around 03 to 00 UTC. The catch, the first outputs he pulled out of it are, by his own account, the worst results he's ever gotten on those prompts. Google has confirmed nothing. There is no release date. But a config file doesn't lie about what Google is wiring up.
What actually leaked, and where
Antigravity is Google's agentic dev tool — its answer to the AI IDE wave. This morning @Lentils80 posted a screenshot of a JSON config block titled tieredModelIds. The flash tier pointed at a model nobody had seen before
Breaking. Gemini 3.6 Flash, ID "gemini-3.6-flash-tiered", appeared in Antigravity a few minutes ago
Look at the neighbors in that config. "flashLite" maps to gemini-3.1-flash-lite. "pro" maps to gemini-3.1-pro-low. Only the middle "flash" slot jumped to a 3.6 ID. My read. Google is mid-transition between the 3.1 and 3.6 generations, and the cheap/fast tier is the first domino to move. The "-tiered" suffix also suggests this build is wired into Antigravity's routing logic rather than exposed as a clean API model — an internal-testing artifact, not a launch announcement.
Worth being precise about what this is, a leaked internal identifier. No blog post, no API listing, no pricing page, no date. Google has said nothing publicly, and everything below the config line is inference.
The first outputs are genuinely rough
About two hours after the first post, @Lentils80 ran the leaked ID on his standard test prompts. He was not kind about the results
Some of the first Gemini 3.6 Flash outputs for y'all, and yeah. Genuinely the worst results I've ever gotten on these two prompts. Terrible frontend capability (2 shot btw) and spatial reasoning, at least it's very fast ig. I cope this isn't actually running on High thinking level but idk
To be fair to the screenshot, the thing it built isn't nothing — a dark-themed voxel world builder with a sun-cycle slider, fog control, particle density settings, and a stats footer showing 126,132 voxels at 61 FPS. A year ago that would've been a decent demo. But the tester's bar is calibrated by current frontier models, and his verdict on frontend capability and spatial reasoning was blunt, worst he's ever gotten on those two prompts, two shots allowed. The one positive he gave it unprompted, it's very fast.
What the leak doesn't tell us — read this before judging
Before anyone writes the "Gemini 3.6 Flash is bad" post, three big caveats
- A two-prompt sample from one tester is not a benchmark. It's an anecdote with screenshots. Useful as a smoke signal, useless as a ranking.
- Early checkpoints ship broken configurations. The tester himself suspects the model wasn't running at High thinking level. Leaked IDs inside routing configs frequently default to the cheapest inference setting — you may be looking at the model's floor, not its ceiling.
- Flash is the trade-off tier. Flash-tier models exist to trade peak quality for speed and cost. Judging one against frontier-class prompts and being shocked it loses is like being surprised a commuter bike loses a drag race. Speed is the feature, the tester noticed it even while panning the output.
One more wrinkle worth flagging, this leak pattern is becoming a genre. Model IDs surface inside vendor tools days or weeks before launch, a tester yanks outputs out of a half-configured build, and the first impression circulates as if it were a review. It happened with Gemini rumor cycles earlier this month — we covered the 3.5 Pro "dragon" chatter the same way. The information in the config file is real. The quality verdict attached to it usually isn't.
My read, the honest takeaway from today is "Google is testing a 3.6 Flash inside its own IDE," full stop. Everything about quality is premature until the model is live, at a known thinking level, on tasks people can reproduce.
What this means for developers choosing models
This is exactly why we don't score models off leaked screenshots. A model earns a seat in your stack when it's live and measurable on real tasks — that's the entire premise of our model benchmarks, where models run the same coding tasks under the same harness so the comparison means something. When Gemini 3.6 Flash ships publicly, that's where its number comes from, not from a config file at 3 a.m.
The competitive context matters more than the leak itself. Qwen and Kimi's open-weight models keep getting better on the exact cheap/fast axis Flash owns — see today's Qwen3.8 Max Preview update for how fast that tier is moving. If Google ships a weak Flash against improving open weights, the bottom of its lineup gets exposed — that's the tier where developers route bulk agent traffic and where price-per-task decides everything.
What to watch for next
- The ID mutating or disappearing in Antigravity configs — Google has pulled leaked IDs within days before.
- A non-tiered "gemini-3.6-flash" identifier showing up in API-adjacent surfaces, which would signal a real launch build.
- More testers running it at High thinking — if outputs jump in quality, today's rough results were a config artifact.
- Any official word from Google. Until then, file this under "leak, unconfirmed," keep building on models you can measure, and check the benchmarks board when 3.6 Flash actually lands.
