Gemini 3.8 Flash and 3.8 Flash Cyber: Features, Benchmarks, Pricing, and What They Mean for AI Agents
Google’s Gemini Flash lineup is moving quickly. Gemini 3.8 Flash, released in September 2026, is designed to deliver stronger coding, reasoning, and agentic capabilities while keeping the low-cost positioning of the Flash family.
Alongside it, Google has introduced Gemini 3.8 Flash Cyber, a specialized variant focused on cybersecurity and vulnerability discovery, with access limited to trusted defenders.
In this article, we’ll look at what Gemini 3.8 Flash offers, how it performs against Gemini 3.7 Flash and other models, what is different about the Cyber variant, how much the models cost, and where they fit into real-world AI agent and coding workflows. |
What Is Gemini 3.8 Flash?
Gemini 3.8 Flash is Google’s latest Flash-tier model, designed for coding, long-horizon software engineering, autonomous agents, and enterprise workflows.
The model sits below Google’s Pro-tier models but is significantly cheaper, making it attractive for developers who need to run AI agents at scale.
Its introductory pricing is:
|
One of the major changes compared with Gemini 3.7 Flash is that Gemini 3.8 Flash is designed to spend more effort on difficult tasks.
It can perform additional reasoning steps and use tools repeatedly, which can improve performance on complex coding and agentic workflows.
The trade-off is that higher effort can consume more tokens.
What's New With Gemini 3.8 Flash?
Gemini 3.8 Flash focuses heavily on developers building AI agents rather than only users looking for a conversational assistant.
1. Adjustable Effort Levels
Developers can adjust how much effort the model uses depending on the task.
Higher effort can provide:
|
Lower effort can provide:
|
This makes the effort setting an important cost-versus-performance control.
For straightforward tasks, maximum reasoning may not be necessary.
For complex software engineering or autonomous workflows, additional reasoning can be worthwhile.
2. Long-Horizon AI Agents
One of the biggest strengths of Gemini 3.8 Flash is its focus on long-running agentic tasks.
Instead of responding to one coding request, an agent can potentially:
|
Google reports strong performance on DeepSWE v1.1, a benchmark designed around long-horizon software engineering.
This is particularly relevant for developers building coding agents that need to work through repositories and complex engineering problems rather than generate isolated code snippets.
3. Stronger Coding and Tool Use
Gemini 3.8 Flash's biggest benchmark improvement is on Terminal-Bench 2.1.
The model achieved:
90.8%
compared with:
81.6% for Gemini 3.7 Flash
Terminal-Bench evaluates how reliably a model can complete command-line and coding tasks from beginning to end, including tool usage and error recovery.
A nearly nine-point improvement suggests that Google has made significant progress in agentic coding and terminal-based workflows.
Gemini 3.8 Flash Benchmarks
The benchmark results are not equally strong across every category.
The largest gains appear in coding, tool use, and agentic workflows, while some traditional reasoning benchmarks remain relatively flat.
Benchmark | Gemini 3.8 Flash | Gemini 3.7 Flash |
Terminal-Bench 2.1 | 90.8% | 81.6% |
SWE-Bench Pro | 61.6% | 60.4% |
SWE-Atlas | 51.9% | 48.0% |
τ³-bench Banking | 38.1% | 30.9% |
CharXiv (multimodal) | 86.2% | 84.5% |
Humanity's Last Exam | 45.4% | 45.7% |
The table reveals something important: Gemini 3.8 Flash is not an across-the-board upgrade of the same magnitude.
Its biggest improvement is in agentic coding and tool-based tasks.
Terminal-Bench 2.1
Gemini 3.8 Flash scores 90.8% on Terminal-Bench 2.1.
For comparison:
Gemini 3.8 Flash — 90.8%
Gemini 3.7 Flash — 81.6%
GPT-5.6 Terra — 87.4%
Claude Sonnet 5 — 80.4%
This makes the result particularly notable for a Flash-tier model.
The benchmark is also closer to the experience of building a coding agent because the model has to execute tasks rather than simply answer coding questions.
DeepSWE v1.1
Gemini 3.8 Flash also performs strongly on DeepSWE v1.1, a long-horizon software engineering benchmark.
Google positions the model as outperforming many larger frontier models while operating at a much lower cost.
The exact percentage is not provided in the supplied benchmark information, but the key point is that Gemini 3.8 Flash is being optimized for agents that can solve complex engineering problems end-to-end.
For developers building autonomous coding systems, this type of benchmark can be more meaningful than a single-shot coding test.
SWE-Bench Pro and SWE-Atlas
The gains are more modest on repository-level software engineering benchmarks.
SWE-Bench Pro
Gemini 3.8 Flash — 61.6%
Gemini 3.7 Flash — 60.4%
SWE-Atlas
Gemini 3.8 Flash — 51.9%
Gemini 3.7 Flash — 48.0%
The SWE-Bench Pro improvement is relatively small.
This reinforces the idea that Gemini 3.8 Flash's biggest advantage isn't simply writing better code. Its stronger gains appear when the model must interact with tools and complete multi-step workflows.
Reasoning Performance Is More Mixed
On HLE-Verified, Gemini 3.8 Flash reaches 54.9%, showing strong multi-step reasoning across STEM, humanities, and professional domains.
However, its performance on the broader Humanity's Last Exam benchmark is essentially unchanged:
Gemini 3.8 Flash — 45.4%
Gemini 3.7 Flash — 45.7%
That is an important caveat.
If your workload primarily involves difficult open-ended reasoning rather than coding, tool use, or autonomous workflows, the upgrade from Gemini 3.7 Flash may not provide the same level of improvement.
Gemini 3.8 Flash Cyber
Alongside the standard model, Google introduced Gemini 3.8 Flash Cyber, a specialized version designed for cybersecurity work.
The Cyber variant is focused on areas such as:
|
Access is restricted to trusted defenders through Google's Fairwind Program.
The underlying intelligence is shared with the standard Gemini 3.8 Flash model, but the Cyber variant is designed around specialized cybersecurity workflows.
Cybersecurity Performance
Google reports that Gemini 3.8 Flash Cyber achieves a real-world vulnerability discovery rate above 70% on its internal evaluation.
It also performs strongly on CWE-Bench, an external benchmark focused on vulnerability patching.
Google reports a 47.2% pass@1 result, placing the model on the benchmark's Pareto frontier while operating at a significantly lower cost than some larger models.
The focus is particularly interesting because the model is positioned toward defensive security, rather than simply maximizing offensive exploitation capabilities.
As with other vendor-reported benchmark results, independent reproduction will be important before treating these numbers as universal measures of cybersecurity performance.
Gemini 3.8 Flash for Enterprise Workflows
The model isn't limited to software development.
Google is also positioning Gemini 3.8 Flash for professional workflows in areas such as:
|
It reportedly improves on benchmarks such as Vals Finance Agent V2 and Harvey's Legal Agent Benchmark.
This matters because enterprise AI agents increasingly need to process large amounts of information and complete multi-step workflows rather than simply generate text.
For example, an agent could potentially:
Read documents → Analyze information → Apply business rules → Generate a report → Verify the result
That type of workflow is where the combination of reasoning and tool use becomes especially valuable.
Gemini 3.8 Flash in AI Studio and Antigravity
Gemini 3.8 Flash is also integrated into Google's developer ecosystem, including Google AI Studio and Antigravity.
The goal is to allow developers to give an agent a high-level instruction and let it iterate through the implementation.
For example:
"Build an interactive dashboard from this CSV, add filters, create three visualizations, and test the application."
Instead of producing only an initial code draft, the agent can potentially continue through the development loop:
Build → Test → Identify errors → Fix → Test again
This is where higher effort levels can become useful because the model has more room to reason and iterate.
Gemini 3.8 Flash Pricing
One of Gemini 3.8 Flash's biggest advantages is its price.
Through December 31, 2026
Pricing | Cost per 1M Tokens |
Input | $0.75 |
Output | $3.75 |
From January 1, 2027
Pricing | Cost per 1M Tokens |
Input | $1.50 |
Output | $7.50 |
This places Gemini 3.8 Flash well below higher-priced frontier models.
For organizations running large numbers of AI-agent tasks, the difference can become significant.
However, token price should not be the only consideration.
A cheaper model that requires many additional iterations can sometimes cost more overall than a more expensive model that completes the task in fewer steps.
The real metric is therefore cost per successful task, not simply cost per million tokens.
Gemini 3.8 Flash vs. Other Models
Gemini 3.8 Flash is particularly interesting because it sits between low-cost AI infrastructure and frontier-level performance.
Model | Best Suited For |
Gemini 3.8 Flash | Coding agents, automation, high-volume agentic workloads |
Gemini 3.7 Flash | Efficiency-first workloads |
Claude Fable 5.1 | Long-running, high-complexity agent tasks |
Claude Sonnet 5 | Everyday coding and general AI work |
GPT-5.6 Terra | Advanced reasoning and coding |
Gemini 3.8 Flash's biggest competitive advantage is the combination of agentic performance and low pricing.
Should You Upgrade From Gemini 3.7 Flash?
It depends on your workload.
Upgrade if you:
|
Stay with Gemini 3.7 Flash if you:
Prioritize low token usage
|
The benchmark results suggest that the upgrade is most valuable for coding and agentic workloads, rather than every type of reasoning task.
What Gemini 3.8 Flash Means for AI Agents
The bigger story behind Gemini 3.8 Flash isn't simply another model release.
It reflects a broader shift toward AI agents that can complete tasks autonomously.
Traditional AI:
Prompt → Response
Agentic AI:
Goal → Plan → Tools → Execution → Testing → Correction → Result
Gemini 3.8 Flash is clearly optimized for the second approach.
Its combination of low pricing, tool use, adjustable reasoning effort, and long-horizon coding capabilities makes it especially interesting for developers who want to run agents at scale.
Conclusion
Gemini 3.8 Flash is a significant step forward for Google's Flash lineup, particularly in coding, terminal tasks, tool use, and autonomous AI agents.
Its 90.8% Terminal-Bench 2.1 score and low introductory price make it a strong option for developers building cost-sensitive agentic applications.
However, the improvements are not universal. Some reasoning benchmarks remain nearly unchanged, so the value of upgrading depends heavily on the workload.
The Gemini 3.8 Flash Cyber variant also shows how AI models are becoming increasingly specialized for professional domains such as cybersecurity.
Overall, Gemini 3.8 Flash's biggest proposition is simple:
Bring stronger agentic AI performance to more developers without the cost of a traditional frontier model.
Comments