Gemini 3.8 Flash Is Already in AI Mode — Three Weeks After 3.7
By Paul Lovell · September 3, 2026 · 5 min read
Google announced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on September 2, 2026, and the general-purpose model is already available inside Google Search's AI Mode.
For search marketers, the headline isn't the benchmark table. It's that the model generating AI Mode answers changed roughly three weeks after the last one did.
What shipped
Two models:
- Gemini 3.8 Flash — the general-purpose model, which Google describes as its "most intelligent workhorse model yet," with improvements across software engineering, agentic tasks and multi-step reasoning.
- Gemini 3.8 Flash Cyber — a cybersecurity-specialised variant for vulnerability discovery and patching, which is not generally available.
Google's framing is that 3.8 Flash "works harder" on difficult tasks by "executing extra reasoning steps, and calling tools iteratively." That phrasing is worth holding onto — more iterative tool calling is the part with implications for how pages get fetched and cited.
The part that matters for search
Gemini 3.8 Flash is live in AI Mode now, but with a constraint that's easy to miss in the coverage. Google's own wording is precise:
"3.8 Flash is available to Google AI Pro and Ultra subscribers across the Gemini app, AI Mode in Google Search and Gemini in Google Sheets."
Paid subscribers only. Free-tier AI Mode is not running 3.8 Flash today.
It's also worth noting what Google didn't say. The announcement doesn't describe 3.8 Flash as replacing the default model in AI Mode — and as of Google's December 2025 AI Mode update, the stated default was Gemini 3 Flash "for AI Mode globally," with other models reached through the AI Mode model drop-down menu. Nothing in this announcement claims that default changed.
So this is not, today, a change that reshapes AI Mode output for the bulk of searchers. If your AI Mode visibility moved this week, this release is unlikely to be the cause at current distribution. That calculus changes if and when access widens beyond paid tiers.
Beyond Search, 3.8 Flash is also in the Gemini app (Pro/Ultra), Gemini in Google Sheets, Google AI Studio, Android Studio, Antigravity, Stitch and Gemini Enterprise.
The benchmarks, including the one Google didn't win
Google's published figures for 3.8 Flash:
- HLE-Verified: 54.9% on multi-step reasoning across STEM, humanities and professional fields
- DeepSWE v1.1: outperforms most larger frontier models on autonomous software engineering
- Vals Finance Agent V2 and Harvey's Legal Agent Benchmark: ahead of 3.7 Flash and other frontier models
- Gray Swan: described as a "significant leap in prompt injection robustness"
For the Cyber variant, Google cites frontier-level performance on CyberGym, an internal vulnerability benchmark success rate above 70% across 20 programming languages, Chrome's Security team finding it produces 2.6x more correct patches than leading commercial models, and Wiz measuring 7.5–9.7% higher recall at 2.3–5.2x lower cost.
One number is worth flagging because it cuts against the launch narrative: on CWE-Bench, Cyber scores a Pass@1 of 47.2%, against a leading competing model's 47.8%. Google published a benchmark where it comes second. That's more candour than these announcements usually contain, and it's a reasonable reminder that "frontier-level" is a range, not a ranking.
As always, these are the vendor's own selected benchmarks and framing. Treat them as directional.
Pricing
$0.75 per million input tokens and $3.75 per million output tokens — the same introductory rate as 3.7 Flash.
That introductory pricing expires on December 31, 2026. From January 1, 2027, it doubles to $1.50/1M input and $7.50/1M output. If you're building SEO tooling on the API, budget for the step change rather than the launch price.
The Cyber variant is gated
Gemini 3.8 Flash Cyber is not openly available. Access runs through Google's Fairwind Program and is restricted to government authorities, critical infrastructure operators and software maintainers, with an application required.
That gating is the story in itself: a model good enough at finding vulnerabilities that Google would rather hand it to defenders under review than ship it broadly.
What to actually do about this
- Stop attributing AI Mode changes to core updates by default. The model behind AI Mode is now iterating on a roughly three-week cycle, independently of ranking updates. A shift in how you're summarised or cited can happen with no algorithm update at all — a distinction worth keeping straight, as we noted after the May 2026 core update landed two days after I/O.
- Measure it natively. Search Console's generative AI performance reporting is the only first-party way to separate AI Mode and AI Overviews visibility from classic organic. Industry datasets tell you what to look for; your own numbers tell you what happened.
- Don't re-plan your content around this release. It's opt-in, paid-tier only, and three weeks from now there may be another model. Optimising for citation — clear answers, original data, well-labelled structure — is durable across model versions in a way that chasing any single release is not. The cited-versus-uncited gap is still the variable you can actually influence.
- If you build on the API, diarise December 31. Introductory pricing ending is a 100% cost increase on both input and output.