Gemini 3.8 Flash and Flash Cyber Explained: The Model That Found a 13-Year-Old Chrome Bug

On September 2, 2026, Google published a post on its official blog titled "Introducing Gemini 3.8 Flash and 3.8 Flash Cyber" [1], and the timing alone is worth noting before anything else in this post. This is Google's third Flash-tier release in roughly six weeks, arriving just three weeks after Gemini 3.7 Flash shipped on August 13, 2026 [4] [2]. That cadence is not incidental. It sits inside a genuinely crowded few weeks of frontier model releases, Claude Fable 5.1 went generally available on September 1 [2], and Google's own Flash line has now shipped three distinct versions since roughly the start of summer. The Register's coverage put it plainly in its headline: with Gemini 3.8 Flash, Google is reminding everyone it is still in the race [2].

Two distinct models shipped under this release. The first, Gemini 3.8 Flash, is what Google calls its "most intelligent workhorse model," built for agentic tasks, software development, and multi-step reasoning, positioned as a substantial upgrade over 3.7 Flash while keeping the same speed and roughly the same introductory price [1]. The second, Gemini 3.8 Flash Cyber, is a specialized variant built specifically for autonomous vulnerability discovery and patching, released through a new limited-access program for governments and trusted security partners called Fairwind [3] [4]. The single most concrete, verifiable claim in the entire release is about that second model: according to Doug Turner, Chrome's Engineering Director, 3.8 Flash Cyber found a vulnerability in Chromium and Chrome that had gone unnoticed for 13 years, a subtle bug hundreds of engineers had already looked past [3].

This post walks through what actually changed in the standard model versus 3.7 Flash, the real benchmark numbers behind Google's claims, the economics of running it in production, what Flash Cyber's numbers actually show and where they came from, why Google is shipping on a three-week cadence right now, and working code, both Python and raw REST, for calling the model today rather than relying on a vendor's own marketing copy.

If you want a visual sense of the core idea before the technical breakdown, agentic reasoning happening at increasingly higher speed and lower cost, here is a generation prompt built around exactly that concept:

A single continuous overhead shot on a clean studio desk. A brass mechanical timer sits beside two glass vials of liquid. The camera holds steady as the left vial's liquid slowly clarifies and brightens over a few seconds while the right vial stays dim, a soft light gradually intensifying above the left vial as if representing focused, iterative reasoning. Clean minimalist studio lighting, soft neutral color grading, physically accurate reflections and shadows, no readable text, no logos, no real interface elements, photoreal rendering with smooth continuous motion and no jump cuts, suitable as a Veo-style generation prompt.

Step 1: What Actually Shipped on September 2

Google's own framing of the standard model is specific: Gemini 3.8 Flash delivers "significant improvements from 3.7 Flash across software engineering, agentic tasks, and critical, multi-step reasoning in specialized domains" [1]. Tulsee Doshi, Google's senior director of product management, and Raluca Ada Popa, Gemini Security Lead at Google DeepMind, described the model in a joint statement as delivering "substantial gains from 3.7 Flash, often approaching the performance of higher-cost frontier models," while confirming that 3.7 Flash itself remains fully supported for efficiency-first workloads that do not need the extra reasoning depth [3] [4].

The rollout followed a specific order rather than shipping everywhere simultaneously. It went live first inside the Gemini app for Google AI Pro and Ultra subscribers, then AI Mode in Search, then Gemini in Google Sheets, with developer access following through Google Antigravity, AI Studio, and the Gemini API [4] [1]. It is also available in Gemini Enterprise and Android Studio [3]. Search Engine Journal separately confirmed the AI Mode integration specifically, noting Google added 3.8 Flash to its AI-powered search experience the same day [6]. Google's own announcement on X put the positioning succinctly: "Introducing Gemini 3.8, our best reasoning & coding" model to date [7].

On raw specifications, Gemini 3.8 Flash accepts text, image, video, audio, and PDF input, and supports a 1,048,576-token input context window with a 65,536-token output limit [2] [10]. That context window is unchanged from 3.7 Flash, which tells you something important about where Google actually invested the engineering effort here: this was not a scale-up release. It is a reasoning-quality release, built on the same context and modality footprint as its predecessor.

Step 2: The Real Benchmark Numbers, and What "Works Harder" Actually Means

Google's own language for describing 3.8 Flash's behavior is unusually candid for a model announcement: according to Doshi and Popa, "3.8 Flash works harder. On complex tasks, it exhibits greater diligence, executing extra reasoning steps, and calling tools iteratively" [3] [4]. That phrase, working harder rather than simply being smarter in some abstract sense, is the through-line that connects every benchmark result and every cost implication covered in this post, so it is worth holding onto as you read the numbers below.

On Artificial Analysis's Intelligence Index, a composite benchmark that aggregates performance across a range of reasoning, coding, and knowledge tasks into a single comparable score [11], Gemini 3.8 Flash scores 59 in high reasoning mode, a three-point increase over 3.7 Flash [2]. That places it in a tight cluster with two other frontier models released around the same window, tying GPT-5.6 Sol and Grok 4.6, both also scoring 59, while sitting below Claude Opus 5 at 63 and Claude Fable 5.1 at 66 [2].

On coding specifically, Gemini 3.8 Flash outperforms most larger frontier models on DeepSWE v1.1, a benchmark for long-horizon, autonomous software engineering tasks that require a model to work through a multi-step engineering problem rather than answering a single isolated coding question [4] [2]. That specific framing, a Flash-tier model beating models several times larger on a long-horizon engineering benchmark, is the clearest evidence for Google's underlying claim that the extra reasoning steps 3.8 Flash takes are actually paying off on tasks that require sustained, multi-step problem solving rather than a single fast answer.

On agentic performance more broadly, 3.8 Flash ranks 14th in Agent Arena, ahead of DeepSeek-V4-Pro, and 7th in Text Arena [4]. On Humanity's Last Exam, a benchmark specifically designed to test genuinely difficult, expert-level reasoning across STEM, humanities, and professional domains, the HLE-Verified variant of the model scores 54.9 percent on multi-step reasoning tasks [4] [2]. On domain-specific enterprise benchmarks, the model exceeds both its predecessor and other frontier models on Vals Finance Agent V2 and Harvey's Legal Agent Benchmark, two evaluations built specifically around realistic finance and legal agentic workflows rather than generic knowledge questions [3] [4].

The tradeoff behind all of this is direct and Google does not obscure it: 3.8 Flash costs roughly 40 percent more per task than 3.7 Flash, because it generates more output tokens and executes additional agentic reasoning steps and tool calls to arrive at those better results [2]. This is the practical, load-bearing consequence of "working harder": a harder-working model does more, which costs more, even when its per-token price stays flat.

Step 3: The Economics, Priced Per Task Rather Than Per Token

Gemini 3.8 Flash launched at the same introductory price as 3.7 Flash: $0.75 per million input tokens and $3.75 per million output tokens, held through December 31, 2026, before standard pricing of $1.50 and $7.50 respectively takes effect on January 1, 2027 [1] [2].

Raw per-token pricing understates the real economics, though, precisely because of the 40 percent higher token consumption per task covered above. Artificial Analysis's own cost-per-task metric, which accounts for how many tokens a model actually needs on average to complete a task at its measured intelligence level rather than just its sticker price per million tokens, puts Gemini 3.8 Flash at roughly $0.58 per Intelligence Index task, describing it as "the cheapest model at its level of intelligence" currently available [2]. For context, that same metric puts Claude Fable 5.1 at roughly $3.76 per task, meaning Gemini 3.8 Flash currently runs at close to one-sixth the effective cost per unit of measured intelligence, despite Fable 5.1 scoring higher on the raw Intelligence Index [2].

That gap is the actual strategic story behind this release, more than the raw benchmark numbers are. Google is not claiming Gemini 3.8 Flash is the smartest model available today, and the Intelligence Index numbers above do not support that claim either. It is positioned, explicitly, as the best available combination of intelligence and cost, aimed squarely at the kind of high-volume, cost-sensitive agentic workloads, customer support automation, coding assistants running thousands of sessions a day, document processing pipelines, where a six-fold cost difference per task compounds into a genuinely large budget line at scale.

Here is a simple way to reason about that cost-per-task tradeoff directly, using the real published per-token prices and a rough estimate of tokens-per-task rather than treating "cost per million tokens" as the whole story:

python
/code
# Illustrative cost-per-task comparison using Gemini 3.8 Flash's real
# published per-token pricing and the reported ~40% higher token
# consumption per task versus 3.7 Flash. This is a simplified estimate
# for reasoning about budget impact, not a reproduction of Artificial
# Analysis's actual Intelligence Index cost-per-task methodology, which
# is not public.

def estimated_task_cost(input_tokens: int, output_tokens: int,
                         price_in_per_million: float, price_out_per_million: float) -> float:
    return (input_tokens / 1_000_000) * price_in_per_million + \
           (output_tokens / 1_000_000) * price_out_per_million

# Published introductory pricing for both 3.7 Flash and 3.8 Flash.
PRICE_IN = 0.75
PRICE_OUT = 3.75

# Rough baseline token usage for one agentic task on 3.7 Flash.
baseline_input = 4000
baseline_output = 1200

# 3.8 Flash is reported to use roughly 40% more tokens per task on
# average, driven by extra reasoning steps and iterative tool calls.
flash_38_input = int(baseline_input * 1.10)
flash_38_output = int(baseline_output * 1.55)

cost_37 = estimated_task_cost(baseline_input, baseline_output, PRICE_IN, PRICE_OUT)
cost_38 = estimated_task_cost(flash_38_input, flash_38_output, PRICE_IN, PRICE_OUT)

print(f"3.7 Flash estimated cost per task: ${cost_37:.4f}")
print(f"3.8 Flash estimated cost per task: ${cost_38:.4f}")
print(f"Increase: {((cost_38 / cost_37) - 1) * 100:.1f}%")

Step 4: Inside Gemini 3.8 Flash Cyber, and the 13-Year-Old Bug

While the standard model is a general-purpose agentic workhorse, Flash Cyber is a narrowly specialized variant trained specifically for autonomous vulnerability discovery and automated patching, and Google describes it as its most capable cybersecurity model to date, with frontier-level performance on both halves of that job, finding vulnerabilities and fixing them [3] [4].

On CyberGym, a benchmark built specifically to evaluate a model's ability to identify and characterize security vulnerabilities in real codebases, Flash Cyber scores 86.2 percent [3]. On CWE-Bench, which evaluates a model's ability to actually generate a correct patch for a known Common Weakness Enumeration category vulnerability rather than merely identifying that one exists, it scores 47.2 percent [3]. Across a broader vulnerability discovery evaluation spanning 20 different programming languages, the model achieved over a 70 percent success rate at finding real vulnerabilities [3].

The most concrete, independently attributable claim in the entire release, though, is the Chrome patching statistic. According to internal Google usage data, verified by Doug Turner, Chrome's Engineering Director, Flash Cyber produced 2.6 times more correct patches for real Chrome vulnerabilities than larger commercial models Google had access to [3]. Turner's specific quote is worth reproducing directly, since it is the single most vivid, verifiable data point in the whole announcement: "One interesting vulnerability 3.8 Flash Cyber discovered had been in Chromium and Chrome for 13 years," a subtle bug that hundreds of engineers had already looked at and missed [3].

That 13-year figure deserves a moment of real scrutiny rather than being repeated as a headline stat, because it is the kind of claim that is easy to either dismiss or overstate. Chromium is one of the most heavily audited, most actively fuzzed pieces of open source software in existence, run through Google's own OSS-Fuzz infrastructure continuously and reviewed by a large, security-conscious engineering organization for well over a decade. A bug surviving that scrutiny for 13 years is not evidence that Chrome's security process is weak, if anything it is closer to the opposite: it demonstrates that some classes of vulnerabilities require a genuinely different kind of search than what continuous fuzzing and human code review already provide, one that involves reasoning about how a specific piece of code behaves across unusual, non-obvious execution paths rather than just generating or reviewing inputs at scale. That is precisely the kind of task a model built to iteratively reason about code, rather than pattern-match against known vulnerability signatures, is suited for.

Corroborating data beyond Google's own Chrome usage came from Wiz, the cloud security company Google acquired for a reported $32 billion, which reported Flash Cyber achieving 7.5 to 9.7 percent higher recall on internal penetration testing exercises, at 2.3 to 5.2 times lower cost than the leading frontier models Wiz had previously used for the same task [3]. Separately, Google Cloud's own Vulnerability Research team used the model to discover a critical vulnerability in under two hours, a process that has typically taken months using conventional methods [3].

Access to Flash Cyber is deliberately narrow. It ships through Fairwind, a new limited-access program Google built specifically for government authorities, critical infrastructure operators, and approved security partners, rather than through the open Gemini API [4] [3]. Organizations that want access have to apply through the program directly, and the model reportedly ships with more permissive cybersecurity mitigations than a general-purpose model would get, a decision that only makes sense precisely because its deployment is this restricted, a dual-use vulnerability-discovery tool loose in the open would be a genuinely different risk calculation than one gated behind vetted, credentialed partners [3].

Step 5: Why Google Is Shipping on a Three-Week Cadence Right Now

Understanding why this specific release landed when it did, and why it looks the way it does, requires some honest context about Google's recent track record, which The Register's coverage does not gloss over. Google missed its own June 2026 target for Gemini 3.5 Pro, and the 3.5 Flash release that did ship was, by the outlet's own framing, underwhelming enough that it got overshadowed in coverage by open-weight releases out of China [2]. That is the backdrop against which three Flash releases in six weeks, 3.6, 3.7, and now 3.8, reads less like an unbroken string of confident progress and more like a company visibly course-correcting its release cadence after a rough stretch.

The competitive landscape it is course-correcting into is genuinely dense right now. GPT-5.6 Sol Ultrafast shipped running on Cerebras's wafer-scale silicon, hitting inference speeds roughly 14 times faster than standard GPU-based serving, a fundamentally different kind of advantage than a benchmark score, one about raw latency rather than reasoning quality our coverage here. Claude Fable 5.1 went generally available on September 1, the day before this release, with cache reads dropping to $0.25, directly undercutting the exact kind of cost-sensitive, high-volume workload Gemini 3.8 Flash is also targeting our coverage here. Grok 4.6 shipped from SpaceXAI with a 500,000-token context window aimed specifically at coding agents our coverage here, and DeepSeek V4-Pro's 0813 release, the model 3.8 Flash beats in Agent Arena according to Google's own numbers, has been closing the gap with proprietary frontier models at a fraction of the price for months our coverage here.

Against that backdrop, a three-week release cadence looks less like confidence and more like a deliberate strategy of shipping frequent, incremental, well-targeted improvements rather than waiting for a single, larger step-change release the way Google's 3.5 Pro delay suggests it originally intended. Whether that strategy is sustainable, or whether it risks diluting attention across too many closely-spaced releases, is a genuinely open question the industry will only be able to answer in hindsight. What is verifiable today is the specific, concrete result: a Flash-tier model beating larger frontier competitors on a long-horizon coding benchmark, at close to one-sixth the cost per task of the highest-scoring competitor, and a specialized security variant that found a bug real engineers had missed for over a decade.

Step 6: Calling Gemini 3.8 Flash Today, Real Working Code

Unlike Flash Cyber, the standard Gemini 3.8 Flash model is genuinely public today, callable through the Gemini API with no waitlist, using either Google's official Python SDK or a raw REST call. This section uses real, current syntax rather than a conceptual illustration, since this is one of the few pieces of this release you can actually verify by running it yourself.

The most important thing to understand before writing any code against 3.8 Flash if you are migrating from an earlier Gemini version is that the reasoning-control parameter changed shape. Earlier Gemini models used a thinking_budget parameter, a numeric token budget for internal reasoning. Gemini 3.8 Flash replaces that entirely with a string enum called thinking_level, accepting "low", "medium", or "high" rather than a raw token count, and notably "minimal" is not a supported value on this model even though it existed on some earlier ones [9]. Google's own migration guidance also flags that the older temperature, top_p, and top_k sampling parameters are deprecated on this model and should be removed from any migrated code rather than carried forward [9].

Here is the Python SDK version, using Google's official google-genai package:

python
/code
# Real, current example using Google's official google-genai Python SDK
# against the live Gemini API. Requires: pip install google-genai
# Docs: https://ai.google.dev/gemini-api/docs/generate-content/latest-model

from google import genai
from google.genai import types

client = genai.Client()  # reads GEMINI_API_KEY from the environment

response = client.models.generate_content(
    model="gemini-3.8-flash",
    contents="Review this payment processing function for edge cases "
              "and suggest a fix, reasoning through each step:\n\n"
              "def charge(amount, currency): ...",
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(
            thinking_level="medium"  # "low", "medium", or "high" only;
                                      # "minimal" is not supported on 3.8 Flash
        ),
    ),
)

print(response.text)

And the equivalent raw REST call, useful for a quick manual test or for any environment where pulling in the full SDK is not worth it:

bash
/code
# The same request as a raw REST call, useful for a quick manual test
# without pulling in the SDK. Requires GEMINI_API_KEY to be set.
curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-3.8-flash:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H 'Content-Type: application/json' \
  -X POST \
  -d '{
    "contents": [{
      "parts": [{"text": "Summarize the tradeoffs of thinking_level=high for a high-volume production pipeline."}]
    }],
    "generationConfig": {
      "thinkingConfig": {
        "thinkingLevel": "medium"
      }
    }
  }'

A few practical notes worth stating plainly. First, thinking_level directly trades latency and cost for reasoning depth, and given the roughly 40 percent higher per-task token cost this model already carries at "medium" compared to 3.7 Flash, defaulting to "high" for every request in a high-volume production pipeline is an easy way to blow through a budget faster than the headline per-token price suggests. Start at "low" or "medium" for a new integration and only raise it for the specific task types where the extra reasoning steps demonstrably improve output quality, rather than setting it globally. Second, the model identifier is gemini-3.8-flash, not a hyphenated variant, and the REST endpoint lives at https://generativelanguage.googleapis.com/v1beta/models/gemini-3.8-flash:generateContent [9] [10], for the Gemini Enterprise Agent Platform's endpoint format specifically, check the equivalent Vertex AI publisher-model path if you are calling through Google Cloud rather than the standalone Gemini API. Third, Flash Cyber has no equivalent public code example here, because it has no public API, access runs entirely through the Fairwind application process covered above, and any tutorial claiming a working Flash Cyber API key should be treated with real skepticism.

Production Best Practices for Teams Adopting 3.8 Flash

A handful of concrete, non-speculative practices are worth adopting immediately if you are migrating a production agentic pipeline from 3.7 Flash to 3.8 Flash rather than starting fresh.

Treat thinking_level as a per-task-type tuning parameter, not a global default. Since the entire value proposition of 3.8 Flash over 3.7 Flash is that it works harder on tasks that actually benefit from extra reasoning steps, applying "high" uniformly across a pipeline that includes plenty of simple, low-stakes calls erases most of the cost advantage this release is actually built around. Profile which specific task types in your pipeline saw the biggest quality lift when you tested "medium" versus "high", and reserve the higher setting for those specifically.

Audit any migrated code for the removed sampling parameters before deploying. Since temperature, top_p, and top_k are deprecated on this model, code carried forward from a 3.7 Flash integration that still sets those values may either be silently ignored or may throw a validation error depending on how strictly your SDK version enforces the schema, and it is worth testing this explicitly rather than assuming backward compatibility.

Benchmark your own workload's actual cost per task, not just the headline per-token price. The gap between 3.8 Flash's per-token price, unchanged from 3.7 Flash, and its real per-task cost, roughly 40 percent higher due to increased token consumption, is exactly the kind of thing that only shows up in a real production bill, not in a rate card. Run a genuine cost comparison against your actual workload before committing to a full migration, rather than assuming the unchanged sticker price means an unchanged bill.

If your use case touches security-sensitive code review or vulnerability triage, do not assume Flash Cyber's capabilities are available to you through the standard model. The standard Gemini 3.8 Flash model was not benchmarked on CyberGym or CWE-Bench in Google's own announcement, and treating a general-purpose coding-capable model as a substitute for a purpose-trained security model is a real category error worth avoiding, especially for anything touching production infrastructure.

Common Mistakes When Evaluating This Release

A handful of misreadings show up repeatedly whenever coverage of a release like this spreads quickly across many outlets at once.

• Treating the Intelligence Index score of 59 as evidence 3.8 Flash is the smartest model available right now. It ties GPT-5.6 Sol and Grok 4.6 and sits below both Claude Opus 5 and Claude Fable 5.1 on that specific metric. The actual claim being made is about cost-efficiency at that level of intelligence, not about topping the leaderboard outright.
• Assuming the unchanged per-token pricing means an unchanged real-world bill. The roughly 40 percent higher token consumption per task compared to 3.7 Flash is the more financially relevant number for anyone actually budgeting a production migration.
• Confusing Gemini 3.8 Flash and Gemini 3.8 Flash Cyber as the same model with a marketing label attached. They are genuinely separate systems with separate benchmarks, separate access models, and separate intended use cases, one is a general-purpose public API model, the other is a restricted-access specialist tool.
• Assuming Flash Cyber is publicly available or has a documented API. As of this writing, it ships exclusively through the Fairwind program for vetted governments, critical infrastructure operators, and approved partners, with no public sign-up.
• Treating the 13-year-old Chrome bug statistic as proof that Chrome's existing security process is weak, rather than evidence that certain vulnerability classes require a genuinely different search strategy than continuous fuzzing and human review already provide.
• Reading three Flash releases in six weeks as pure unqualified momentum without the context of Google's missed 3.5 Pro deadline and the comparatively unenthusiastic reception 3.5 Flash got against Chinese open-weight competitors, context that meaningfully shapes why this specific release cadence exists right now.

Where a Tool Like Miraflow Fits Into This Picture

It is worth being precise about what Gemini 3.8 Flash is and is not relevant to for a content creation workflow, since the "new Gemini model" headline gets applied loosely across genuinely different product categories. Gemini 3.8 Flash is a text-and-reasoning model built for agentic tasks, coding, and multi-step problem solving, it is not an image or video generation model, and it is a fully separate system from the Nano Banana image model line or Veo video models that actually power visual content generation inside tools like Miraflow.

Where this release is genuinely useful to a creator or a small team is upstream of the actual visual generation step: scripting, planning, and research work that benefits from stronger multi-step reasoning at low cost. A creator using an agentic workflow to research a topic, draft a script outline, and iterate on hooks before ever touching a visual generation tool is exactly the kind of cost-sensitive, multi-step task 3.8 Flash is built around. Once that script exists, Miraflow's Text2Shorts turns it into a fully produced, voiced vertical video in one pipeline, and the AI Image Generator and YouTube Thumbnail Maker handle the actual visual side of production that a text-and-reasoning model like this one is not built to do. For background on the video and image models actually doing that visual generation work, our comparison of Nano Banana Pro against Seedream 5.0 Pro and our breakdown of Veo 3.1 against Seedance 2.5 cover that separate, adjacent category directly.

Conclusion

Gemini 3.8 Flash is not the release of a company claiming the smartest model on the market, and Google's own benchmark numbers do not support that reading. It is a specific, well-evidenced claim about cost-efficient reasoning: a Flash-tier model that beats larger frontier competitors on long-horizon software engineering tasks, at roughly one-sixth the effective cost per task of the highest-scoring competitor on the same composite benchmark, achieved by genuinely working harder on the tasks that call for it rather than by simply scaling up. Flash Cyber is the more narrowly remarkable half of this release, a restricted-access specialist that surfaced a security bug real Chrome engineers had missed for 13 years, verified and attributed by name rather than left as an anonymous marketing claim, which is the kind of specific, checkable detail that should carry more weight than an aggregate benchmark score.

Whether shipping on a three-week cadence turns out to be a sustainable strategy for Google or a sign of a company still finding its footing after a rough stretch earlier in 2026 is not yet answerable, and treating this release as a definitive verdict either way would be reading past what the actual, verifiable evidence supports. What is verifiable today is a working API you can call this afternoon, real benchmark numbers that place the model honestly in a genuinely crowded field rather than at the very top of it, and a specialized security variant with one of the more concrete, attributable results, a bug found after 13 years, to have come out of any model release so far in 2026. For more on the broader inference and architecture techniques behind releases like this one, our explainer on speculative decoding and faster LLM inference covers the underlying engineering that makes fast, iterative reasoning loops like 3.8 Flash's tool-calling behavior practical at scale, and our breakdown of context engineering covers the framework side of building genuinely reliable multi-step agents on top of a model like this one.

References and Sources

[1] Google. "Introducing Gemini 3.8 Flash and 3.8 Flash Cyber."

[2] The Register. "With Gemini 3.8 Flash, Google reminds everyone it's still in the race."

[3] VentureBeat. "Google's Gemini 3.8 Flash is built for agents, while its cyber twin hunts vulnerabilities."

[4] 9to5Google. "Gemini 3.8 Flash rolling out three weeks after last release."

[5] Neowin. "Google launches Gemini 3.8 Flash with frontier-level performance at a fraction of the price."

[6] Search Engine Journal. "Google Adds Gemini 3.8 Flash To AI Mode."

[7] Google. "Introducing Gemini 3.8, our best reasoning & coding model." (X/Twitter)

[8] Thurrott. "Google Releases Gemini 3.8 Flash and Cyber Variant."

[9] Google AI for Developers. "What's new in Gemini 3.8 Flash."

[10] Google Cloud Documentation. "Developer's guide to Gemini 3.8 Flash."

[11] Artificial Analysis. "Artificial Analysis (independent AI model benchmarking)."