# Google 发布 Gemini 3.8 Flash，六周内第三款 Flash 模型

- 来源：The Decoder：AI News（RSS）
- 作者：Matthias Bastian
- 发布时间：2026-09-03 00:59
- AIHOT 分数：74
- AIHOT 链接：https://aihot.virxact.com/items/cmtkcfz8s019trog0x5emk0f6
- 原文链接：https://the-decoder.com/gemini-3-8-flash-is-googles-third-budget-model-in-six-weeks-while-frontier-models-remain-mia

## AI 摘要

Google 发布 Gemini 3.8 Flash，含通用版和面向经审核防御者的 3.8 Flash Cyber。

## 正文

Three weeks after Gemini 3.7 Flash, Google is back with Gemini 3.8 Flash. The budget model is supposed to beat its predecessor by a wide margin in coding.

Google is releasing Gemini 3.8 Flash in two versions: a general-purpose reasoning and coding model and a specialized cybersecurity version called 3.8 Flash Cyber. This is the third Flash release in just six weeks, and whether that rapid cadence is a sign of strength or a distraction from the still-missing frontier models Gemini 3.5 Pro and Gemini 4 depends on who you ask. New Deepmind head Koray Kavukcuoglu made it clear that Google isn't just chasing price-performance optimization and still wants to lead on raw capability too.

According to Google's own numbers, Gemini 3.8 Flash hits 73.7 percent on the DeepSWE v1.1 benchmark for long-horizon software engineering tasks. That's just below Claude Opus 5 at 74.0 percent but well ahead of Claude Sonnet 5 (53.8%), GPT-5.6 Sol (72.7%), and the previous 3.7 Flash (65.3%).

Google says Gemini 3.8 Flash beats Anthropic's Opus 5 and OpenAI's GPT-5.6 Sol across many benchmarks despite being far cheaper. These are just benchmarks, though, and real-world performance can feel very different. | Image: Google

Google also says the new model has improved at 3D generation. The video below shows a 3D game that Gemini 3.8 Flash reportedly built from a single prompt in Google's AI coding tool Antigravity. The textures were generated with Google's Nano Banana image model.

视频 · 前往原文观看

Gemini 3.8 Flash launches at a reduced price of $0.75 per million input tokens and $3.75 per million output tokens, matching 3.7 Flash. Starting January 2027, the regular price will rise to $1.50 and $7.50. Claude Opus 5 costs $5.00 per million input tokens and $25.00 for output tokens, while GPT-5.6 Sol sits at $4.00 and $20.00. Even after the introductory pricing expires, Gemini 3.8 Flash would remain far cheaper on a per-token basis than the top models from OpenAI and Anthropic.

But Google explains the performance gains over 3.7 partly by saying that 3.8 Flash runs extra reasoning steps on complex tasks and calls tools iteratively. The model "works harder," Google says, which means higher token consumption that partly offsets the lower per-token price. For use cases where compute efficiency matters most, Google recommends lower reasoning levels or sticking with the still-supported 3.7 Flash.

Gemini 3.8 Flash is available to developers through Google AI Studio, Google Antigravity, and Android Studio. Businesses can access it through Gemini Enterprise. Consumers get it in the Gemini app, in Google Search's AI Mode, and for paying subscribers, in Google Sheets.

Strong bang for the buck, but token costs add up

Independent benchmarking platform Artificial Analysis gives Gemini 3.8 Flash an Intelligence Index score of 59, three points above its predecessor 3.7 Flash at 56. TThat puts it on par with GPT-5.6 Sol at xhigh reasoning and Grok 4.6 at medium reasoning, both also scoring 59. The Intelligence Index gains come mainly from stronger performance on agentic benchmarks like tool use and coding tasks, according to Artificial Analysis.

Gemini 3.8 Flash scores 59 on the Artificial Analysis Intelligence Index, putting it on par with GPT-5.6 Sol and Grok 4.6. On the cost-performance chart (bottom), the model sits on the Pareto frontier, offering the lowest cost per task at its intelligence level. | Image: Artificial Analysis

On cost per task, 3.8 Flash hits the Pareto frontier according to Artificial Analysis, coming in at $0.58 per task as the cheapest model at its intelligence level. But that figure has risen about 40 percent compared to 3.7 Flash at $0.40, even though the per-token price stayed the same. That's likely why Google recommends sticking with 3.7 Flash for efficiency-focused workloads.

At high reasoning levels, 3.8 Flash produces about 300 output tokens per second with an average time per task of 2.5 minutes. That's slightly faster than GPT-5.6 Luna (2.6 minutes) and GPT-5.6 Terra (3.3 minutes) but slower than Claude Fable 5.1 (2.1 minutes) and the older 3.7 Flash (2.2 minutes). At low reasoning levels, the time drops to about 48 seconds.

Cybersecurity model stays locked down for vetted defenders

Gemini 3.8 Flash Cyber, like its predecessor 3.5 Flash Cyber, isn't publicly available. Google distributes it through the Fairwind Program to government agencies, critical infrastructure operators, and software maintainers. The model has less restrictive safety settings than the standard version because it's built for defensive cybersecurity work.

On CyberGym, the industry-standard benchmark for detecting vulnerabilities in C/C++, 3.8 Flash Cyber scores 86.2 percent according to Google, beating the previous 3.5 Flash Cyber (77.5%), GPT-5.6 Sol (83.6%), and GPT-5.5-Cyber (85.6%). For automated patching on the external CWE-Bench, 3.8 Flash Cyber hits 47.2 percent Pass@1, nearly matching the leading frontier model at 47.8 percent while costing much less.

The model also appears more resilient against prompt injection attacks. On the Gray Swan IPI benchmark, Gemini 3.8 Flash achieves an attack success rate of just 5.5 percent. DeepSeek V4 Pro comes in at 60.1 percent, Kimi K3 at 52.7 percent, and Grok 4.6 at 51.8 percent. Only Anthropic's Claude Opus 5 does slightly better at 4.8 percent, and Opus 5 with its additional security options scores even lower within the Claude ecosystem.

Google Blog

AA