Google pushes Gemini 3.8 Flash into the budget AI race

Google is releasing Gemini 3.8 Flash in standard and cybersecurity versions, with stronger coding results and low per-token pricing. The tradeoff is that the model may use more tokens on harder tasks, which can raise the real cost per task.

Google pushes Gemini 3.8 Flash into the budget AI race

Google is moving quickly in its lower-cost AI lineup. Gemini 3.8 Flash arrives only three weeks after Gemini 3.7 Flash, making it the third Flash release in just six weeks.

The new model is aimed at reasoning, coding and practical tool use, while a separate Gemini 3.8 Flash Cyber version is designed for defensive cybersecurity work. The launch strengthens Google's price-performance story, even as attention remains on the still-missing frontier models Gemini 3.5 Pro and Gemini 4.

A fast update cycle for Gemini Flash

Gemini 3.8 Flash is being released in two forms. One is a general-purpose reasoning and coding model. The other is 3.8 Flash Cyber, a specialized cybersecurity model.

The pace matters because Google has now delivered three Flash releases in just six weeks. That can be read in two ways. It shows rapid iteration in budget models, but it also keeps the spotlight on Google's higher-end roadmap while Gemini 3.5 Pro and Gemini 4 remain absent.

New Deepmind head Koray Kavukcuoglu has signaled that Google is not only focused on price-performance optimization. The company still wants to compete on raw capability as well.

For developers and businesses, the practical question is less about release cadence and more about workload fit. Gemini 3.8 Flash is positioned as a model that can handle harder coding and agentic tasks without moving into the price range of top-tier models from OpenAI and Anthropic.

Coding gains are the headline

Google's own numbers put Gemini 3.8 Flash at 73.7 percent on DeepSWE v1.1, a benchmark for long-horizon software engineering tasks. That places it just below Claude Opus 5 at 74.0 percent.

The same comparison shows Gemini 3.8 Flash ahead of Claude Sonnet 5 at 53.8%, GPT-5.6 Sol at 72.7%, and Gemini 3.7 Flash at 65.3%. On that measure, the jump from 3.7 Flash is the clearest upgrade in the release.

Google also says the model has improved at 3D generation. In one example, Gemini 3.8 Flash reportedly created a 3D game from a single prompt in Google's AI coding tool Antigravity. The textures in that example were generated with Google's Nano Banana image model.

The broader pattern is that Gemini 3.8 Flash is being framed as a practical builder's model. Coding, tool use and iterative work appear central to the release, rather than chat alone.

Low token prices do not tell the whole cost story

Gemini 3.8 Flash launches at $0.75 per million input tokens and $3.75 per million output tokens. That matches Gemini 3.7 Flash. Starting January 2027, the regular price rises to $1.50 and $7.50.

Those numbers remain well below Claude Opus 5, which costs $5.00 per million input tokens and $25.00 for output tokens. GPT-5.6 Sol is listed at $4.00 and $20.00. Even after the introductory price ends, Gemini 3.8 Flash would still be cheaper per token than those top models.

But per-token pricing is only part of the picture. Google says some of the performance improvement comes from the model taking extra reasoning steps on complex work and using tools repeatedly. In Google's phrasing, the model "works harder."

That matters because harder work can mean higher token consumption. A model can be cheaper per token and still cost more per completed task if it generates or processes more tokens along the way.

Google's own guidance reflects that tradeoff. For workloads where compute efficiency is the priority, it recommends lower reasoning levels or the still-supported 3.7 Flash.

Independent benchmarks show the same tradeoff

Artificial Analysis gives Gemini 3.8 Flash an Intelligence Index score of 59. That is three points above Gemini 3.7 Flash, which scores 56.

That score puts Gemini 3.8 Flash level with GPT-5.6 Sol at xhigh reasoning and Grok 4.6 at medium reasoning, both also scoring 59. According to Artificial Analysis, the improvement is mainly tied to agentic benchmarks such as tool use and coding tasks.

Cost per task gives a more nuanced view. Artificial Analysis places Gemini 3.8 Flash on the Pareto frontier at $0.58 per task, describing it as the cheapest model at its intelligence level. But that figure is about 40 percent higher than Gemini 3.7 Flash at $0.40, even though the per-token price is unchanged.

Speed also depends on reasoning level. At high reasoning levels, Gemini 3.8 Flash produces about 300 output tokens per second and averages 2.5 minutes per task. That is slightly faster than GPT-5.6 Luna at 2.6 minutes and GPT-5.6 Terra at 3.3 minutes, but slower than Claude Fable 5.1 at 2.1 minutes and Gemini 3.7 Flash at 2.2 minutes.

At low reasoning levels, the time drops to about 48 seconds. That gives users a lever: higher reasoning may improve difficult results, while lower reasoning may keep tasks faster and cheaper.

Gemini 3.8 Flash Cyber stays restricted

The cybersecurity version is not being released publicly. Like its predecessor 3.5 Flash Cyber, Gemini 3.8 Flash Cyber is distributed through the Fairwind Program to government agencies, critical infrastructure operators and software maintainers.

Google gives the Cyber model less restrictive safety settings because it is built for defensive cybersecurity work. That narrower distribution reflects the model's intended use case: helping vetted defenders find and fix security problems.

On CyberGym, described as the industry-standard benchmark for detecting vulnerabilities in C/C++, Gemini 3.8 Flash Cyber scores 86.2 percent according to Google. That beats 3.5 Flash Cyber at 77.5%, GPT-5.6 Sol at 83.6%, and GPT-5.5-Cyber at 85.6%.

For automated patching on the external CWE-Bench, Gemini 3.8 Flash Cyber reaches 47.2 percent Pass@1. That nearly matches the leading frontier model at 47.8 percent while costing much less.

The standard Gemini 3.8 Flash also appears stronger against prompt injection attacks. On the Gray Swan IPI benchmark, it has an attack success rate of 5.5 percent. DeepSeek V4 Pro is listed at 60.1 percent, Kimi K3 at 52.7 percent, and Grok 4.6 at 51.8 percent. Claude Opus 5 is slightly lower at 4.8 percent, with Opus 5 using additional security options scoring even lower within the Claude ecosystem.

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