Gemini 3.8 Flash: Google's New Workhorse Model and the Cyber Variant (2026)

Introduction

Google DeepMind is shipping Flash models at a startup's cadence. Gemini 3.8 Flash arrived on September 2, 2026 — just three weeks after 3.7 Flash, and the third Flash release in six weeks. Alongside it, Google quietly introduced something new: Gemini 3.8 Flash Cyber, a defense-specialized variant that may be the most consequential security model released to date, if you can get access to it.

This guide covers what's new in 3.8 Flash, the pricing cliff coming in January 2027, and why the Cyber variant's restricted release tells you a lot about where AI-assisted security is heading. For the broader assistant, see our Gemini review.

What's New in Gemini 3.8 Flash

Smarter Where It Counts

Google positions 3.8 Flash as its "most intelligent workhorse model," with the upgrades concentrated in exactly the areas developers use Flash for:

• Software engineering — stronger on long-horizon SWE tasks; on DeepSWE v1.1 it outperforms most larger frontier models
• Agentic tasks — more reliable multi-step execution with tool use
• Professional reasoning — HLE-Verified score of 54.9%, covering STEM, humanities, and professional domains

One behavioral change deserves attention: 3.8 Flash is more diligent by default. On complex tasks it takes extra reasoning steps and iterates on tool calls — which improves results but can consume more tokens. If you're running high-volume workloads, use the new lower effort levels to cap token spend.

Availability

Gemini 3.8 Flash is broadly available:

Channel	Access
Gemini API (Google AI Studio)	✅
Google Antigravity, Stitch, Android Studio	✅
Gemini Enterprise	✅
Gemini app, Search AI Mode, Sheets	AI Pro / Ultra subscribers
Chatbot free tier	Rolling out

Pricing: The January 2027 Cliff

This is the most important commercial fact about 3.8 Flash:

Period	Input / 1M	Output / 1M
Now – Dec 31, 2026 (intro)	$0.75	$3.75
From Jan 1, 2027	$1.50	$7.50

The introductory rates match 3.7 Flash exactly, and they double on January 1, 2027. If you're building on Flash — or comparing it against competitors — plan against the $1.50/$7.50 steady-state price, not the intro rate. For context: even at the post-intro price, 3.8 Flash remains dramatically cheaper than frontier models like GPT-6 Astra at $10/$50, and it undercuts Claude Fable 5.1 by more than 13x on input tokens.

Google is also offering 50% off introductory API rates for the Flash family through the end of 2026, so effective costs for eligible developers are even lower right now.

Gemini 3.8 Flash Cyber: The Restricted Sibling

Flash Cyber shares 3.8 Flash's foundational intelligence but is tuned specifically for vulnerability detection and automated patching — defense, not offense. Google states it deliberately prioritized defensive capability over exploit development, which is why access is restricted.

The Evidence

Google published unusually concrete third-party-adjacent numbers:

Test	Result
Internal synthetic vuln discovery (20 languages)	>70% success rate
CWE-Bench patching (pass@1)	47.2% (leading frontier model: 47.8%)
Chrome Security team: correct patches	2.6x more than the best larger commercial model
Wiz pentest benchmark	Recall +7.5–9.7% at 2.3–5.2x lower cost
Google Cloud vuln research	Found a critical vulnerability in under 2 hours (typically months)

Who Can Get It

Flash Cyber is not publicly available. Access runs through the Fairwind Program, limited to:

• Government authorized agencies
• Critical infrastructure operators
• Software maintainers

There is no public pricing — approved organizations work directly with Google. If you're a security team without Fairwind access, the closest openly available alternatives are GPT-6 Astra (restricted to defensive use at the Critical threshold) and Claude Fable 5.1 (vulnerability discovery permitted, exploitation blocked).

3.8 Flash in the 2026 Landscape

Model	Input/Output per 1M	Sweet Spot
Gemini 3.8 Flash	$0.75/$3.75 (intro)	High-volume coding, agents, everyday work
GPT-5.6 Terra	$2.00/$12.00	Balanced professional work (OpenAI)
GPT-6 Astra	$10.00/$50.00	Frontier agentic + computer use
Claude Fable 5.1	$10.00/$50.00	Cache-heavy agent loops, coding

3.8 Flash doesn't compete with the frontier — it competes for the workloads most teams actually run. Against Gemini vs GPT-5.6 at the assistant level, 3.8 Flash strengthens Google's hand on price-sensitive coding while Astra and Fable 5.1 fight it out at the top.

Summary

Gemini 3.8 Flash is an incremental-but-real upgrade over 3.7 Flash: better at software engineering, agents, and professional reasoning, at the same introductory price — with a doubling scheduled for January 2027 that developers should plan around now. The Flash Cyber variant is a landmark for AI-assisted defense, but its Fairwind-only access keeps it out of most hands.

Practical recommendations:

• Building high-volume coding or agentic products → 3.8 Flash at intro pricing is the best value in Western frontier labs' lineups; budget for the 2027 price change
• Security teams → Apply to Fairwind if eligible; otherwise evaluate Astra's defensive capabilities
• Maximum capability regardless of cost → Astra or Fable 5.1 remain the frontier