The new Gemini 3.8 Flash and Gemini 3.8 Flash Cyber: Pricing, Coding, Agents, and Fairwind

Google released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on September 2, 2026, extending the Gemini Flash line in two different directions: a broadly available agentic model for coding and professional reasoning, and a restricted cybersecurity variant designed for trusted defenders.

The distinction is operational rather than cosmetic, because Gemini 3.8 Flash is generally available through Google’s developer and enterprise channels, while Flash Cyber is distributed through the limited-access Fairwind Program under tighter eligibility and safety controls.

For developers, the central trade-off is that 3.8 Flash is designed to spend additional reasoning effort and call tools iteratively on complex tasks, which can improve reliability on long-horizon work while increasing token consumption compared with efficiency-first configurations or earlier Flash models.

Google lists an introductory Gemini 3.8 Flash API price of $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026; Google states that pricing will rise to $1.50 and $7.50 respectively from January 1, 2027.

··········

GEMINI 3.8 FLASH IS A GENERAL-PURPOSE AGENTIC MODEL, WHILE FLASH CYBER IS A RESTRICTED DEFENSIVE SYSTEM.

The two releases share foundational intelligence, but their availability, safeguards, pricing model, and intended operating environment differ sharply.

Gemini 3.8 Flash is positioned as the primary high-capability Flash workhorse for software engineering, multi-step reasoning and autonomous agent workflows, with general availability documented by Google Cloud on the day of release.

........

Product or program	Availability	Primary role	Pricing or access condition
Gemini 3.8 Flash	GA; Gemini API, Google AI Studio, Android Studio, Antigravity, Gemini Enterprise, selected consumer surfaces	Coding, reasoning, long-horizon agents, multimodal professional work	$0.75/M input and $3.75/M output through Dec. 31, 2026; $1.50/M and $7.50/M from Jan. 1, 2027
Gemini 3.8 Flash Cyber	Restricted	Vulnerability discovery, defensive cybersecurity analysis, automated patching	No general public access; prioritized for trusted defenders through Fairwind
Fairwind Program	Limited-access program	Operational layer combining Flash Cyber with CodeMender and Google Cloud security tooling	Available to selected governments, critical-infrastructure operators, software maintainers, Google Cloud customers and security partners

........

Google also keeps Gemini 3.7 Flash available for workloads where compute efficiency is the dominant constraint, which is important because 3.8 Flash can consume additional tokens as it performs more reasoning steps on difficult tasks.

Google Cloud documents LOW, MEDIUM and HIGH thinking levels for 3.8 Flash, with MEDIUM as the default, while MINIMAL is not supported; developers therefore have an explicit control for trading additional inference effort against cost and latency.

··········

CODING AND LONG-HORIZON AGENTS ARE THE MAIN PERFORMANCE TARGET.

Google’s launch data emphasizes end-to-end software engineering and iterative tool use rather than simple single-turn benchmark speed.

Google describes Gemini 3.8 Flash as more diligent than 3.7 Flash on complex tasks, meaning that the model may execute additional reasoning steps and invoke tools repeatedly before producing a result, especially at higher effort levels.

That design choice is relevant for repository-scale coding, autonomous debugging, multi-stage research and professional workflows where an agent has to inspect intermediate results, decide what to do next and continue until a task is complete.

On HLE-Verified, Google reports a score of 54.9% for Gemini 3.8 Flash, while the launch material also claims substantial gains over 3.7 Flash on DeepSWE v1.1 and on specialized finance and legal agent benchmarks.

These are vendor-reported benchmark results and should be interpreted as evidence of Google’s intended positioning rather than as independent confirmation of real-world superiority across every workload.

For production systems, the practical test is whether the extra reasoning and tool calls reduce expensive task failures enough to compensate for higher token use, because an apparently inexpensive per-token model can become materially more costly when autonomous loops run for many steps.

··········

FLASH CYBER MOVES GEMINI FROM VULNERABILITY ANALYSIS TOWARD AUTONOMOUS PATCHING.

The cybersecurity release is built around finding, verifying and fixing vulnerabilities, with access deliberately constrained because its defensive capability set is more permissive than the safeguards applied to the general model.

Google says Gemini 3.8 Flash Cyber was trained and evaluated for vulnerability discovery across complex codebases and for generating patches that can be validated before deployment, while Fairwind combines the model with the CodeMender harness inside controlled cloud environments.

........

Evaluation or deployment claim	Google-reported result	What it indicates	Evidence status
Internal vulnerability benchmark across 20 programming languages	Success rate above 70%	Breadth of autonomous vulnerability discovery beyond C/C++-focused tests	Google internal benchmark
CWE-Bench automated patching	47.2% pass@1 versus 47.8% for a leading frontier model	Competitive patch generation at lower stated operating cost	External benchmark result reported by Google
Chrome Security evaluation	2.6× more correct patches than the best larger commercial models tested	Patch quality on Chrome vulnerability work	Google internal deployment claim
Wiz penetration-testing benchmark	+7.5–9.7% recall at 2.3–5.2× lower cost	Security detection performance and cost efficiency	Partner-reported result published by Google
Google Cloud Vulnerability Research	Critical foundational vulnerability found in under two hours	Potential acceleration of high-complexity defensive research	Google real-world case claim

........

The access model is central to the release: Flash Cyber is not a normal Gemini app option or an unrestricted API model, and Google says Fairwind prioritizes governments, critical-infrastructure operators, software maintainers and other trusted defenders.

Google also states that Flash Cyber uses more permissive cybersecurity mitigations than the general 3.8 Flash model, which explains why the company is separating defensive access from the standard public deployment path.

For organizations that qualify, the operational proposition is a loop in which vulnerabilities can be discovered, fixes generated, patches validated and deployment-ready changes produced inside the organization’s cloud environment rather than treating the model as a standalone code assistant.

··········

THE RELEASE CHANGES THE FLASH TRADE-OFF MORE THAN IT CHANGES THE GEMINI PRODUCT HIERARCHY.

Gemini 3.8 Flash is the stronger choice when autonomous coding, multi-step professional reasoning and repeated tool use justify additional inference effort, while efficiency-first workloads can still favor 3.7 Flash or lower-cost models when the marginal accuracy gain is not economically useful.

The introductory API price makes 3.8 Flash particularly aggressive during 2026, but long-lived production deployments should model the announced January 2027 price increase rather than evaluating total cost from the launch price alone.

Flash Cyber should be evaluated separately because it is a controlled defensive product with different safeguards, access conditions and operational objectives, so its benchmark results do not imply that the public Gemini model has the same unrestricted cybersecurity capabilities.

The practical decision rule is therefore workload-specific: move agentic coding and reasoning tasks to 3.8 Flash when higher task completion is worth additional token use, retain efficiency-oriented models where throughput dominates, and treat Fairwind as a specialized security deployment for eligible organizations rather than as a consumer Gemini feature.

··········

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

[datastudios.org]