# Gemini 3.8 Flash Arrives With Frontier-Level Results at Flash Prices

> Google has launched Gemini 3.8 Flash, a faster and comparatively inexpensive model aimed at coding, autonomous agents, and complex professional work. Alongside it comes Gemini 3.8 Flash Cyber, a restricted model designed to help trusted defenders discover and patch software vulnerabilities at scale.
>
> OpenCheese AI · 2026.09.02 · Web page: https://opencheese.dev/blog/gemini-3-8-flash-arrives-with-frontier-level-results-at-flash-prices

Google has introduced two new additions to its rapidly expanding AI lineup: **Gemini 3.8 Flash** and **Gemini 3.8 Flash Cyber**.

The release comes only three weeks after Gemini 3.7 Flash. At this rate, model version numbers may soon need their own live dashboard.

Despite the quick turnaround, this is more than a cosmetic update. According to [Google’s announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/), Gemini 3.8 Flash delivers substantial improvements in long-horizon software engineering, autonomous agents, tool use, and complex reasoning across specialized professional fields.

Its companion, Gemini 3.8 Flash Cyber, applies the same emphasis on speed and agentic execution to vulnerability discovery and automated patching.

## A Flash model that competes with frontier models

Google describes Gemini 3.8 Flash as its most intelligent “workhorse” model so far. It is designed for applications that need more reasoning than a lightweight model typically provides but cannot justify frontier-model prices for every request.

The company’s published benchmarks show especially strong results in financial analysis, legal workflows, terminal coding, chart reasoning, video understanding, and scientific research.

Here are several highlights from Google’s comparison:

| Benchmark                      | Gemini 3.8 Flash | Gemini 3.7 Flash | Claude Opus 5 | GPT-5.6 Sol |
| ------------------------------ | ---------------: | ---------------: | ------------: | ----------: |
| DeepSWE v1.1                   |            73.7% |            65.3% |         74.0% |       72.7% |
| Vals Finance Agent v2          |        **61.4%** |            59.0% |         58.6% |       53.8% |
| Harvey’s Legal Agent Benchmark |        **10.0%** |             8.8% |          6.7% |        2.5% |
| Terminal-bench 2.1             |        **89.4%** |            85.8% |         89.1% |       88.8% |
| CharXiv Reasoning              |        **86.2%** |            84.5% |         83.7% |       85.8% |
| HLE-Verified                   |        **54.9%** |            53.6% |         54.4% |       54.5% |

On long-horizon software engineering, Gemini 3.8 Flash reaches 73.7% on DeepSWE v1.1—just behind Claude Opus 5 at 74.0% and ahead of GPT-5.6 Sol at 72.7%.

It takes the lead in Google’s table on the finance and legal benchmarks, narrowly tops both models on Terminal-bench 2.1, and also performs strongly in chart reasoning and multidisciplinary expert reasoning.

The biology results are notable as well. Gemini 3.8 Flash scores 56.5% on the “human difficult” portion of BioMysteryBench, compared with 49.4% for Claude Opus 5 and 44.7% for GPT-5.6 Sol. It also reaches 86.2% on LABBench2, ahead of the other models included in Google’s comparison.

## Strong results, but not a universal winner

The benchmark table does not show Gemini 3.8 Flash winning everything—and that is worth acknowledging before the victory parade blocks traffic.

Claude Opus 5 remains substantially ahead on Terminal-bench 4.0, GDPval-AA v2, and OSWorld 2.0. GPT-5.6 Sol leads on the GDP.PDF document-comprehension benchmark.

The more accurate conclusion is that Gemini 3.8 Flash is unusually competitive for its price category. It approaches or surpasses considerably more expensive models in several important areas, while still leaving room for specialized frontier models to lead on other tasks.

These are also vendor-published evaluations rather than independent real-world comparisons. Benchmarks are useful signals, but they are not a substitute for testing a model against your own repositories, prompts, tools, and failure cases.

## The pricing may be the real headline

Gemini 3.8 Flash keeps the introductory pricing previously offered for Gemini 3.7 Flash:

* **$0.75 per million input tokens**
* **$3.75 per million output tokens**

That introductory offer runs through **December 31, 2026**. From January 1, 2027, pricing is scheduled to increase to $1.50 per million input tokens and $7.50 per million output tokens.

Even at the regular price, Gemini 3.8 Flash remains below the prices listed for several larger competitors in Google’s comparison. During the introductory period, the difference is difficult to ignore: Claude Opus 5 is listed at $5 per million input tokens and $25 per million output tokens, while GPT-5.6 Sol is listed at $4 and $20 respectively.

There is an important detail hidden behind the inexpensive token price. Google says 3.8 Flash “works harder” on complex requests by performing additional reasoning and calling tools iteratively. That may produce better results, but it can also consume more tokens at higher effort levels.

Developers prioritizing predictable efficiency can choose a lower reasoning effort or continue using Gemini 3.7 Flash. In other words, cheap tokens are still capable of arriving in large groups.

## A million-token context window

According to Google’s [Gemini API documentation](https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash), the model supports:

* A **1,048,576-token context window**
* Up to **65,536 output tokens**
* Text, image, video, audio, and PDF inputs
* Text output
* Low, medium, and high thinking levels

The large context window and multimodal inputs make the model suitable for repository-scale coding tasks, document analysis, long videos, research collections, and agentic workflows that need to maintain substantial working context.

## Gemini 3.8 Flash Cyber focuses on defense

The second model, **Gemini 3.8 Flash Cyber**, is not a general-purpose public release. It is Google’s most capable cybersecurity model and is being made available to trusted defenders through the company’s new **Fairwind Program**.

Its primary jobs are autonomous vulnerability discovery, verification, and patch generation.

Google reports that the model achieves frontier-level performance on CyberGym and exceeds a 70% success rate on an internal vulnerability-discovery benchmark spanning complex codebases and 20 programming languages.

On CWE-Bench, an external patching evaluation, Gemini 3.8 Flash Cyber reaches a 47.2% pass rate, close to the 47.8% result of a leading frontier model but at a significantly lower cost.

Google says it is already using the model internally:

* The Chrome Security team reportedly received 2.6 times more correct vulnerability patches than from the best larger commercial models it tested.
* Wiz measured 7.5 - 9.7% higher recall on an internal penetration-testing benchmark at 2.3–5.2 times lower cost.
* Google’s Cloud Vulnerability Research team used the model to identify a critical foundational vulnerability in under two hours—a task that would typically require months of research.

Because advanced vulnerability discovery is inherently dual use, Gemini 3.8 Flash Cyber has more permissive cybersecurity capabilities than the general model and is limited to approved government authorities, critical-infrastructure operators, software maintainers, and other trusted defenders.

## Where Gemini 3.8 Flash is available

Developers can access Gemini 3.8 Flash through:

* [OpenCheese.Dev](https://opencheese.dev/en/)
* Google Antigravity
* The Gemini API and Google AI Studio
* Android Studio
* Stitch

Enterprise customers can use it through Gemini Enterprise. Google AI Pro and Ultra subscribers can access it across the Gemini app, AI Mode in Google Search, and Gemini in Google Sheets.

## The bottom line

Gemini 3.8 Flash does not replace every frontier model, but it makes the economics of using one harder to defend for many workloads.

It nearly matches or beats significantly more expensive models on several coding, professional, multimodal, and scientific benchmarks. At the same time, Gemini 3.8 Flash Cyber shows how specialized models may become practical defensive tools rather than expensive research demonstrations.

The most interesting part of this release is not that Google won a few benchmark rows. It is that “Flash” increasingly describes the price and speed, not the intelligence ceiling.

## Related posts

- 2026.10.03 · [Google Introduces Gemini 4 Argon With a 1 Million-Token Output Limit](https://opencheese.dev/blog/google-introduces-gemini-4-argon-with-1-million-token-output-limit.md): Gemini 4 Argon is built for long, multi-step coding, business and cybersecurity work. Google is first testing it with trusted cyber defenders, with broader access planned after further safety work.
- 2026.09.04 · [GPT‑6 Astra: OpenAI’s New Frontier for Agents, Coding and Computer Use](https://opencheese.dev/blog/gpt-6-astra-openai-s-new-frontier-for-agents-coding-and-computer-use.md): GPT‑6 Astra is OpenAI’s new flagship model for autonomous computer use, software engineering, science and professional workflows. It posts major benchmark gains, but its unprecedented cybersecurity capabilities also require stronger safeguards.
- 2026.09.29 · [Claude Sonnet 5.5 Arrives With Faster Coding and Stronger Everyday Performance](https://opencheese.dev/blog/claude-sonnet-5-5-arrives-with-faster-coding-and-stronger-everyday-performance.md): Anthropic’s latest Sonnet model brings faster responses, stronger coding results and a large context window. Here is what the early numbers show.