Google officially released Gemini 3.8 Flash and the dedicated cybersecurity version 3.8 Flash Cyber on September 2. This is the third Flash update in six weeks, following 3.5 → 3.6 → 3.7 → 3.8. Google has been updating the Flash line roughly every three weeks.
The official positioning is straightforward: 3.8 Flash is the “smartest workhorse model,” with major improvements in multi-step reasoning for software engineering, agent tasks, and professional domains. More notably, Google has released a dedicated cybersecurity version at the same time—the first such release among mainstream AI models.
Benchmarks: Flash pricing, flagship performance
First, here is Google’s full published benchmark table:
Here are the key results in plain English:
| What is being measured | 3.8 Flash | 3.7 Flash | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Long-horizon software engineering | 71.0% | 65.3% | 74.0% | 72.7% |
| Knowledge work composite | 1545 | 1482 | 1824 | 1710 |
| Financial analysis agent | 61.4% | 59.0% | 58.6% | 53.8% |
| Legal workflow | 10.0% | 8.8% | 6.7% | 2.5% |
| Terminal coding (2.1) | 89.4% | 85.8% | 89.1% | 88.8% |
| Terminal coding (4.0) | 19.1% | 11.2% | 51.8% | 37.3% |
| Chart information understanding | 86.2% | 84.5% | 83.7% | 85.8% |
| Long-video understanding | 87.8% | 85.4% | 75.4% | 82.1% |
| Expert reasoning | 54.9% | 53.6% | 54.4% | 54.5% |
| Computer use | 59.0% | 50.6% | 75.4% | 62.6% |
| Biological research (hard problems) | 56.5% | 43.5% | 49.4% | 44.7% |
| Laboratory research | 86.2% | 82.1% | 84.2% | 82.1% |
Where 3.8 Flash leads: financial analysis, legal agents, long-video understanding, difficult biological research, laboratory research, chart understanding, and Terminal-bench 2.1. These are all practical, domain-focused tasks.
Where it still loses: Terminal-bench 4.0 (19.1% vs. Opus 5’s 51.8%, a huge gap), OSWorld computer use (59.0% vs. 75.4%), and the knowledge work composite (1545 vs. 1824).
In plain terms: for finance, law, biology, video, and charts, 3.8 Flash is no longer behind flagship models. But if you want AI to write software autonomously for long periods or operate a computer, Opus 5 is still far ahead.
How much better is it than 3.7 Flash?
| What is being measured | 3.7 → 3.8 | Gain |
|---|---|---|
| Biological research (hard problems) | 43.5% → 56.5% | +13.0 |
| Computer use | 50.6% → 59.0% | +8.4 |
| Terminal-bench 4.0 | 11.2% → 19.1% | +7.9 |
| Long-horizon software engineering | 65.3% → 71.0% | +5.7 |
| Terminal coding (2.1) | 85.8% → 89.4% | +3.6 |
| Long-video understanding | 85.4% → 87.8% | +2.4 |
The biggest gains in three weeks are on difficult biological research (+13 percentage points) and computer use (+8.4). The official announcement says that training for cybersecurity capabilities also improved general coding and reasoning performance.
One more notable point: 3.8 Flash also shows a “significant improvement” in defending against malicious prompt injection, where someone deliberately attacks the model with instructions, according to Gray Swan test data.
Gemini 3.8 Flash Cyber: dedicated cybersecurity version
The most unusual part of this release is Google’s simultaneous launch of 3.8 Flash Cyber, a version specifically for cybersecurity defenders. It uses the same underlying model as the standard version but has more relaxed safety restrictions, allowing security professionals to detect and remediate vulnerabilities.
Cyber version: real-world results
Google published several highly persuasive figures:
- Chrome security team: When fixing vulnerabilities with 3.8 Flash Cyber, the correct-fix rate was 2.6 times higher than the best commercial model.
- Wiz penetration testing: Detection was 7.5–9.7% higher than flagship models, while the cost was only 1/2.3 to 1/5.2 as much.
- Google Cloud vulnerability research: Found a critical vulnerability in under 2 hours; research of this kind previously typically took months.
- CWE-Bench remediation: 47.2% pass@1, close to the leading model’s 47.8%, but at far lower cost.
Who can use it
3.8 Flash Cyber is not open to everyone. Google has established the Fairwind Program, which is limited to:
- Trusted government agencies
- Critical infrastructure operators
- Software maintainers
Applicants must pass a review before they can use it. This follows the same idea as Anthropic’s Mythos / Cyber Verification Program: give defenders stronger tools while restricting who can access them.
Same price, promotional pricing through year-end
It is exactly the same price as 3.7 Flash:
| Effective period | Input / per million tokens | Output / per million tokens |
|---|---|---|
| Through 2026-12-31 | $0.75 | $3.75 |
| From 2027-01-01 | $1.50 | $7.50 |
For comparison: Claude Opus 5 costs $5/$25, while GPT-5.6 Sol costs $4/$20. 3.8 Flash costs about one-seventh as much as Opus 5. Even after doubling to $1.50/$7.50 in 2027, it remains less than half the price of its competitors.
This is promotional pricing through the end of the year. For long-term budgets, use the 2027 price of $1.50/$7.50 directly.
Where to use it
| Audience | Channels |
|---|---|
| Developers | Google Antigravity, Google AI Studio, Android Studio, Stitch (UI generation), Gemini API |
| Businesses | Gemini Enterprise |
| Individuals | Gemini app (Google AI Pro / Ultra subscriptions), Google Search AI Mode, Google Sheets |
| Security teams | Apply for 3.8 Flash Cyber through the Fairwind Program |
Specifications at a glance
| Item | Details |
|---|---|
| Model ID | gemini-3.8-flash |
| Positioning | The smartest workhorse Flash model |
| Context window | 1 million tokens |
| Maximum output | 64K tokens |
| Thinking levels | low / medium (default) / high |
| Input formats | Text, images, video, audio, PDF |
| Output format | Text |
| Safety features | CBRN protection, enhanced prompt-injection defense |
Little Penguin’s takeaways
- Flash-tier pricing has entered flagship territory. Finance, legal workflows, biological research, and long-video understanding—areas where only flagship models used to perform well—now deliver similar or better results at one-seventh the price. Google’s three Flash updates in six weeks are compressing the room competitors have for a pricing advantage.
- The Terminal-bench 4.0 gap (19.1% vs. 51.8%) tells us something important. Having AI write complex software autonomously over long, multi-step sessions is still flagship territory. Flash’s “smartness” is concentrated in understanding and judgment, not autonomous execution.
- 3.8 Flash Cyber is a signal. Google has turned “AI cybersecurity” from an add-on feature into a standalone product line. The Chrome team’s 2.6× correct-fix rate and Wiz’s cost data are concrete results. If your company works in security, the Fairwind Program is worth watching.
- Three updates in six weeks also create pressure. For developers, switching model versions every three weeks carries meaningful integration and testing costs. Google’s rapid iteration benefits its own ecosystem (Antigravity users upgrade automatically), but API teams need to consider a version-pinning strategy.