> ## Content Index
> Fetch the complete content index at: https://www.metatalks.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Google ships its third Flash model in six weeks, and a security twin it will not sell openly
- URL: https://www.metatalks.ai/google-gemini-3-8-flash-holds-the-old-price-but-uses-more-tokens/
- Published: 2026-09-03T09:19:00.000Z
- Updated: 2026-09-03T09:18:59.000Z
- Author: Al
- Tags: News, Frontier Models, AI Security, #newswire

**The rate per token is the one 3.7 Flash launched with, but the model is built to spend more of them on hard problems — so what a given task costs can move even though the price list has not.**

Google DeepMind has released [Gemini 3.8 Flash](https://deepmind.google/blog/introducing-gemini-3-8-flash-and-38-flash-cyber/?ref=metatalks.ai) at the rates 3.7 Flash carried when it launched: $0.75 per million input tokens and $3.75 per million output tokens. It arrives three weeks after 3.7 Flash.

On DeepSWE v1.1, a software engineering test run over long horizons, Google puts 3.8 Flash ahead of most bigger frontier models at solving engineering problems end to end, and at a fraction of their cost.

Those gains trace back to a design decision. Faced with a hard problem, the model puts in more effort, running additional reasoning steps and making repeated calls to tools, and it may spend extra tokens doing so, particularly when the effort setting is high.

It also makes a clean cost comparison between the two models harder to draw. Where compute efficiency is the main constraint, Google points developers to a lower effort setting, or back to 3.7 Flash, which it continues to support in full.

A second release, Gemini 3.8 Flash Cyber, is pointed at finding software vulnerabilities and repairing them. Google calls it the strongest security model it has built and says repair was the priority set at the outset, ahead of attack-side abilities such as exploitation. On [CWE-Bench](https://cwe-bench.com/?ref=metatalks.ai#leaderboard), an outside test of patching ability administered by Collinear, it puts the model’s pass@1 at 47.2% against 47.8% for a leading frontier model, at a far smaller price.

That model ships with looser restrictions than 3.8 Flash, so Google is handing it only to defenders it trusts: government authorities, operators of critical infrastructure and software maintainers get priority through a new program it calls Fairwind.