共有:

Gemini 3.8 Flash

Gemini's Flash tier moves to 3.8
——same price, heavier under the hood

On September 2, 2026, Google rolled out general availability for "Gemini 3.8 Flash," its lightweight model. It's built on the same foundation as its predecessor, 3.7 Flash, but accuracy and reliability have both been pushed up a notch. The price didn't move — so why did the workload get heavier?

AI Navigate Editorial2026.09.056 min read

GEMINI FLASH LINEAGE 3.5 Flash-Lite 3.6 Flash 3.7 Flash Default thru 9/1 3.8 Flash Current since 9/2 Refreshed roughly once a year, price held flat
01
Why Now

Why Google swaps out Flash
once a year without a price hike

The lightweight-model market is where per-token pricing gets fought over hardest.

Google shipped "Gemini 3.8 Flash" to general availability on September 2, 2026, replacing "Gemini 3.7 Flash" as the default lightweight tier. Google DeepMind's official Flash model page shows the lineage running 3.5 Flash-Lite → 3.6 Flash → 3.7 Flash → 3.8 Flash — a refresh cadence of roughly once a year.

What's notable is that Google didn't raise the price for this refresh at all. The lightweight, low-cost tier is a market where every vendor is competing head-to-head on per-token rates, so holding price flat while pushing capability up looks like the only option Google had left. The tight, roughly year-long release cadence itself is a sign of how intense that competitive pressure has become.

02
By The Numbers

Specs and pricing:
what changed, what didn't

Line up the model-card numbers and the "box" is identical — only what's happening inside it changed.

PRICE VS. TOKENS PER REPLY 3.7 Flash 3.8 Flash Same price ($0.75 / $3.75 per 1M tokens) Tokens consumed per reply trend upward
FIG. The per-token price holds flat while tokens consumed per reply trend upward (conceptual).
1,048,576
input context window (tokens)
65,536
max output tokens
$0.75
per 1M input tokens
$3.75
per 1M output tokens

According to Google DeepMind's Gemini 3.8 Flash model card, the input context window is 1,048,576 tokens and the max output is 65,536 tokens — neither figure has changed from 3.7 Flash. Supported input remains multimodal (text, image, audio, video, PDF) with text-only output, the same setup as before. Pricing was also held flat, at $0.75 per 1M input tokens and $3.75 per 1M output tokens.

What changed isn't the shape of the box, but its behavior inside. Google describes 3.8 Flash as more accurate and more reliable than 3.7 Flash, while noting that this comes at the cost of higher token consumption per response. To let developers dial that trade-off up or down, Google also added a new control called "thinking level" — worth noting in its own right.

03
Who It Helps

Who benefits,
and who won't notice

How much this matters depends heavily on how you already use Flash.

Engineers running Flash at volume

For high-frequency API workloads, accuracy and reliability go up while the per-token rate stays the same. But since output tokens per response can rise, monthly cost needs to be re-modeled as "rate × tokens × calls," not just "rate × calls."

Everyday business users

For people who don't think about which model they're on and just use Gemini's default tier, this is a pure upgrade — no switching required, slightly smarter answers for the same effort. Day-to-day drafting, summarizing, and document work all get a small accuracy bump.

04
Before / After

3.7 Flash and 3.8 Flash,
side by side

The numbers alone look modest, but lined up they make the shape of the change clear: same box, different behavior inside it.

3.7 Flash (through 9/1)3.8 Flash (from 9/2)
Input context: 1,048,576 tokensUnchanged — same 1,048,576 tokens
Max output: 65,536 tokensUnchanged — same 65,536 tokens
$0.75 in / $3.75 out (per 1M tokens)Held flat — identical pricing
Lightweight-tier baseline accuracyHigher accuracy/reliability (more tokens per reply)
Response depth: fixed default onlyAdjustable via thinking level

Same sticker price doesn't mean the same total bill.

05
What To Do Next

Three things to check
before you switch

In many environments, the default model has likely already flipped to 3.8 Flash automatically. Don't let that slide by unchecked.

01

Check your thinking level default

The accuracy-vs-token trade-off hinges on this setting. Before shipping to production, choose it explicitly based on the latency and cost you can tolerate, rather than trusting whatever the default happens to be.

02

Measure token consumption on high-volume workloads

The per-token rate hasn't moved, but longer responses mean higher real cost. For the first week or two after switching, track tokens consumed alongside call volume, not just call volume alone.

03

Plan your timeline off of 3.7 Flash

Google hasn't published an end-of-life date for 3.7 Flash. But given the pace of past generational swaps, it's reasonable to assume the older generation will eventually be deprecated at some point.

06
Counterpoint

A more sober read:
"basically noise"

That said, not everyone has reason to celebrate this refresh. For people who prioritized raw speed above all else and deliberately kept using 3.7 Flash for it, the accuracy bump is hard to feel — and the honest read is that if you were juggling versions for speed, the difference is basically noise. Public information doesn't point to any major improvement on the speed metric itself.

On top of that, the trade-off between higher accuracy and higher token consumption is hard to size up fully from published model-card numbers alone. The fact that the sticker price is unchanged doesn't mean total cost is unchanged — and that gap widens the more heavily a setup calls on Flash. Whether or not you get the upside without the downside will likely come down to how well you use thinking level.