Gemini 3.8 Flash: Faster Than I Expected to Like
After three bad Flash releases in a row, 3.8 is finally good. Fast, cooperative, included with Google AI Ultra, and still no 3.5 Pro while Flash is on its fourth release. 6.5 out of 10.
Gemini 3.8 Flash surprised me, and I do not say that about many models. It comes included with my Google AI Ultra subscription, which means I tried it mostly because it was already there. I now use it on purpose.
Three bad ones first
It is worth remembering how we got here, because 3.8 did not arrive out of nowhere.
Gemini 3.5 Flash was abhorrent, and the pricing made it worse. Around three dollars per million tokens in, for a model with Flash in the name. Flash is supposed to mean cheap and quick. They charged frontier prices for something that was neither, at a time when the alternatives were already better and cheaper.
3.6 Flash fixed the pricing. The model was still shit.
3.7 Flash arrived and the model was still shit.
Then 3.8, which is genuinely good. Three releases of nothing and then the one that works. I have no idea what changed and I am not going to pretend otherwise, but the difference is not subtle.
It is fast in a way that changes how you use it
Snappy is the word. The answer is in front of you before you have finished thinking about the question, and that speed changes what you are willing to ask. Slow models make you batch your requests and think hard before spending one. Flash does not, so you iterate instead of planning.
What I actually tested it on
I used it to stand up demo marketplaces, throwaway environments where I can exercise a new integration against the PCIFIC tools without touching anything real. That is unglamorous work with a lot of moving parts, and it is exactly where a model tends to fall over.
It did it absolutely fine. No arguing, no wandering off, no half-built environment that needed an hour of repair before I could use it.
The front end is the weak spot
Left to its own judgement, the design it produces is not good. Give it skills to work from and it is perfectly fine, which tells you the limitation is taste rather than capability. It can build the thing. It cannot decide what the thing should look like.
That is a fair trade at this speed, as long as you know to bring the design direction yourself.
Where is 3.5 Pro?
The strange part is that 3.5 Pro still does not exist. Flash has shipped four times, all the way from 3.5 to 3.8, while Pro sits where it has sat since 3.1. They have iterated hard on the cheap end and left the top of the range alone, which is not the order I would have done it in.
3.1 Pro is still an insanely smart model, so it is not as though the ceiling is embarrassing. It is that the version numbers have drifted so far apart that Flash and Pro no longer sound like the same product line.
Whenever the next Pro does land, being Google, it will be called something like 3.4 Ultra Pro. They will find a way to make the number go backwards and add two words to it. I have stopped trying to predict the naming.
The verdict: 6.5/10
I rate Gemini 3.8 Flash 6.5 out of 10.
It is fast, it understands what you asked, and it does the boring setup work without a fight. It is not going to design anything for you and it is not the model I hand hard problems to. For scaffolding an environment so I can get on with the real work, it earns its place, and it costs me nothing extra to keep around. After 3.5, 3.6 and 3.7, an alright model is a bigger jump than the number suggests.