Gemini 3.8 Flash shipped three weeks after 3.7. The price tag didn't move, the bill did
Google's third Flash release in six weeks keeps the same per-token price as 3.7 Flash, unlike the cut that came before it. Read past the sticker and a real task now runs roughly 40% more expensive anyway, for the same reason 3.7 Flash quietly got pricier than it looked. There's also a second model in this release, gated to a program most readers have never heard of, built to find security vulnerabilities on its own.
What shipped: Gemini 3.8 Flash, generally available 2 September 2026, three weeks after 3.7 Flash, plus a second model, Gemini 3.8 Flash Cyber, built specifically for vulnerability discovery and patching. Google calls the pair its "next-generation intelligence for agentic workflows and cybersecurity."
The sticker price held, the real cost didn't. Input/output pricing stays at $0.75 / $3.75 per million tokens through 31 December 2026, unchanged from 3.7 Flash, still doubling to $1.50 / $7.50 on 1 January 2027. But Artificial Analysis measured roughly 30% more output tokens per task on its benchmark suite, pushing the real cost per completed task up about 40%, from $0.40 to $0.58.
3.8 Flash Cyber isn't something you can sign up for. It's restricted to Google's Fairwind Program: government cyber-defense authorities, critical-infrastructure operators, and open-source maintainers. There's no public API, no self-serve access, no listed price.
Gemini Notebook still isn't mentioned. Same as with 3.7 Flash, Google's launch materials list the Gemini app, AI Studio, Antigravity, Android Studio, Gemini Enterprise, and the Gemini API. Gemini Notebook does not appear.
The numbers, and which ones are whose
Google's own benchmark table, published alongside the model, shows real gains: HLE-Verified climbing to 54.9%, CWE-Bench pass@1 at 47.2%. Independent measurement from Artificial Analysis puts the model's Intelligence Index at 59 with high reasoning effort, up from 3.7 Flash's 56, and a 12-point jump on the τ³-Banking agentic benchmark to 45%. Those two sources agree closely enough to trust.
A third figure is worth flagging rather than repeating as fact: DataCamp's write-up cites Terminal-Bench 2.1 at 90.8% and SWE-Bench Pro at 61.6%, both notably higher than anything in Google's own post or in 9to5Google's coverage. Neither Google nor Artificial Analysis publishes a matching number. Treat that pair as one outlet's own testing, not a confirmed Google benchmark, until a second independent source reproduces it.
Metric 3.7 Flash 3.8 Flash
--------------------------------------------------------
HLE (plain) 45.7% 45.4%
HLE-Verified n/a 54.9%
CWE-Bench pass@1 n/a 47.2%
AA Intelligence Index (high) 56 59
tau3-Banking (agentic) 33% 45%
Cost per Intelligence Index task $0.40 $0.58
Plain HLE barely moved. The gains concentrate in agentic
and verified-reasoning tasks, not raw knowledge recall,
and they cost more tokens to reach.Why the flat price tag is misleading
This is the second release in a row where the headline price number undersells what a task actually costs. When 3.6 Flash became 3.7 Flash in August, the sticker price was cut in half, but independent testing found the model's heavier default reasoning tier added roughly 40% more billed tokens per task, eating most of that cut for high-volume, low-complexity work. This time there's no cut to erode: 3.8 Flash's per-token price is identical to 3.7 Flash's, and Artificial Analysis's own measurement shows the same pattern repeating anyway. The model now uses about 30% more output tokens to reach a given benchmark score, and thinking tokens still bill at the output rate. Net effect: a task that cost $0.40 on 3.7 Flash's Intelligence Index suite now costs roughly $0.58 on 3.8 Flash, a 40%-plus increase with a sticker price that never moved. If your workload is high-volume and cost-sensitive rather than reasoning-heavy, that's the number to budget against, not the per-token rate.
3.8 Flash Cyber: a model you probably can't use
The second model in this release gets far less coverage than the price story, and it's arguably the more unusual product decision. Gemini 3.8 Flash Cyber is a variant tuned specifically for security work: Google's own figures claim a real-world vulnerability discovery rate above 70%, 2.6x more correct patches for a set of Chrome bugs than commercial rival tools in Google's testing, and a case where it found a critical vulnerability in under two hours, work Google says normally takes a security team months. On Wiz's benchmark it scored 7.5 to 9.7 percentage points higher recall at 2.3 to 5.2 times lower cost than the tools Wiz compared it against.
None of that is something most readers can go try. Cyber ships only through Google's Fairwind Program, restricted to government cyber-defense authorities, operators of critical infrastructure, and maintainers of widely used open-source software. There's no public API endpoint, no consumer or developer signup, and no listed price. If you're evaluating AI coding or security tools for your own team, 3.8 Flash Cyber isn't one of the options on the table, regardless of how the benchmarks read.
Where the regular model actually ships
Standard 3.8 Flash is live in the Gemini app (including Gemini Spark, the agentic assistant, for Google AI Pro and Ultra subscribers), Google Sheets' AI features on the same paid tiers, Google Antigravity, AI Studio, Android Studio, Stitch, the Gemini Enterprise Agent Platform, and the Gemini API. That's the same surface list 3.7 Flash shipped to three weeks earlier, which tells you this is a swap-in model update for Google's existing Flash-tier surfaces rather than a new integration push.
The fatigue is showing, not just in comments
The Register's own coverage of this launch leads with skepticism rather than the benchmark table, calling this Google's third Flash release in six weeks and its fourth in four months, and noting the pattern arrives during a period when Gemini 3.5 Pro, the flagship model, remains delayed well past its original mid-2026 target. Threads discussing the release on Hacker News echo the same read: real complaints about "still no Pro model," and a sense, as one outlet put it, that this is "a treadmill, not a staircase." None of that changes what shipped, but it's worth knowing you're not the only one wondering whether a third Flash refresh in six weeks with no matching Pro release is a cadence or a stall.
What still doesn't connect to Gemini Notebook
Same finding as with 3.7 Flash, worth restating rather than assuming carried over: nothing in Google's announcement, the model card, or the coverage of this launch mentions Gemini Notebook. The confirmed footprint is the general Gemini app, Sheets, developer surfaces, and enterprise tools, not the notebook product specifically. Our own version-history guide tracks which Gemini model actually powers Gemini Notebook's chat and Studio outputs over time, and as of today that guide's most recent confirmed entry still predates this release. If a faster or cheaper Flash model reaching your notebook chats would change how you use it, that hasn't been confirmed, in either direction.
People also ask
Is Gemini 3.8 Flash cheaper than 3.7 Flash?
The per-token price is identical: $0.75 input / $3.75 output per million tokens through 31 December 2026, rising to $1.50 / $7.50 on 1 January 2027. But independent testing found the model uses roughly 30% more output tokens per task, pushing real cost per completed task up about 40% despite the flat sticker price.
What is Gemini 3.8 Flash Cyber?
A security-focused variant of 3.8 Flash built for vulnerability discovery and patching, restricted to Google's Fairwind Program (government cyber-defense authorities, critical-infrastructure operators, and open-source maintainers). It has no public API, no self-serve signup, and no listed price.
Does Gemini 3.8 Flash power Gemini Notebook?
Not confirmed. Google's launch materials list the Gemini app, Sheets, AI Studio, Antigravity, Android Studio, Gemini Enterprise, and the Gemini API as where it ships. Gemini Notebook isn't named anywhere in the announcement, the same gap as the 3.7 Flash release three weeks earlier.
Why did Google ship another Flash model so soon after 3.7?
Google frames it as an accelerated release cadence for its Flash line, its third release in six weeks and fourth in four months. It arrives while Gemini 3.5 Pro, the flagship model promised for mid-2026, remains delayed; Google hasn't linked the two directly, but outlets including The Register have noted the pattern.
Is Gemini 3.8 Flash actually better than 3.7 Flash?
On verified and agentic benchmarks, yes by a real margin: Artificial Analysis measured its Intelligence Index at 59 versus 56, and Google's own figures show gains on HLE-Verified and CWE-Bench. Plain factual-recall performance (HLE) barely moved, so the improvement is concentrated in reasoning and agentic tasks specifically, not general knowledge.