Google launches Gemini 3.8 Flash and a cyber variant

The New Models

Google introduced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber in a company blog post dated 2 September 2026, describing 3.8 as its best reasoning and coding model so far at the same speed and low cost as the earlier 3.7 Flash.

Google said the release is its third Flash model in six weeks and follows 3.7 Flash by three weeks. Both new variants are built on the same foundational intelligence, which the company says was refined by long-running agentic loops that repeatedly evaluate and improve the underlying models, along with training in cybersecurity.

On the benchmark measures Google cited, 3.8 Flash scored 54.9% on HLE-Verified, a test of multi-step reasoning across science, humanities and professional subjects. Google also said it outperforms most larger frontier models on DeepSWE v1.1, a long-horizon software engineering test, and beats 3.7 Flash on Vals Finance Agent V2 and Harvey's Legal Agent Benchmark.

Cost And Behaviour

Google attributes the gains to a design choice it summarises as the model working harder, taking extra reasoning steps and calling tools repeatedly. That can mean higher token use, particularly at higher effort settings.

For developers who care most about compute efficiency, Google says they can select lower effort levels or stay on 3.7 Flash, which remains fully supported. The blog post does not give a price per million tokens for 3.8 Flash.

As demonstrations, Google said 3.8 Flash produced a 3D game in which the player is a wizard navigating a castle, with textures generated by its Nano Banana tool, and a working DOS-style version of Google Maps from a single prompt. Both were built in Google Antigravity.

The Fairwind Program

The Hacker News reported that Gemini 3.8 Flash Cyber, which Google calls its most capable cybersecurity model, is being released to selected defenders through a new initiative called the Fairwind Program. Google said the programme gives early access to governments, healthcare providers and telecommunications services so they can build defences before new threats arrive.

According to the same report, Google is working with more than 650 partners worldwide, including CrowdStrike, Datadog, Menlo Security, Palo Alto Networks and Snowflake, and access extends to some Google Cloud customers and government agencies. The model arrives a little over a month after Gemini 3.5 Flash Cyber, and The Hacker News said it surpasses Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol and GPT-5.5-Cyber at autonomously finding vulnerabilities.

Tulsee Doshi, senior director of product management, and Raluca Ada Popa, Gemini Security Lead at Google DeepMind, said the company invested in fixing vulnerabilities from the start and prioritised that over offensive capabilities such as exploitation. How organisations apply to Fairwind, and when the cyber model might reach a wider audience, is not stated in the sources.

Wider Context

The launches coincided with Anthropic releasing Claude Fable 5.1 and Claude Mythos 5.1 with different levels of safeguards, the latter only through trusted access programmes covering cybersecurity and the life sciences, The Hacker News reported.

Google's own recap of August 2026 said the Gemini app passed 1 billion monthly users during that month, alongside the Pixel 11 series and the launch of 3.7 Flash at an introductory price of half the original 3.6 Flash cost per million tokens.