|

Google DeepMind Releases Gemini 3.8 Flash and Gemini 3.8 Flash Cyber: One Core Model, Two Access Envelopes

ByRicardo

Google simply announced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, three weeks after Gemini 3.7 Flash and marking the third Flash launch in six weeks. Both variants run on the identical foundational intelligence, refined by means of long-running agentic loops that recursively consider the underlying fashions. What separates them isn’t structure, it’s the security envelope and who’s allowed by means of it.

Is it deployable? Gemini 3.8 Flash is mostly accessible as we speak by means of the Gemini API, Google AI Studio, Antigravity, Android Studio, and Gemini Enterprise, so you may route manufacturing visitors to it now. Weights are closed, so there is no such thing as a self-hosted or on-premises path. Gemini 3.8 Flash Cyber isn’t overtly deployable in any respect: entry is granted case by case by means of the brand new Fairwind Program.

What really modified in 3.8 Flash

The research team explains that 3.8 Flash is predicated on 3.7 Flash. Specs are unchanged: a 1,048,576-token context window, 65,536-token most output, textual content, picture, audio, and video enter, and textual content output. Thinking ranges stay LOW, MEDIUM, and HIGH with MEDIUM because the default. One breaking element for anybody migrating: MINIMAL isn’t supported on 3.8 Flash and setting it returns an API validation error.

The behavioral change is the purpose. Google describes the acquire bluntly: 3.8 Flash works tougher. On advanced duties it executes further reasoning steps and calls instruments iteratively, and it could burn extra tokens at greater effort ranges. Google’s personal developer guide is candid that this buys higher accuracy at the price of greater token consumption, and it recommends staying on 3.7 Flash when compute effectivity is the binding constraint. That is an unusually direct admission that the newer mannequin isn’t the proper default for each workload.

https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/

Where the features present up

On DeepSWE v1.1, a long-horizon software program engineering benchmark, Google studies 3.8 Flash outperforming most bigger frontier fashions at a fraction of the fee. It data 54.9% on HLE-Verified, and features over 3.7 Flash and different frontier fashions on Vals Finance Agent V2 and Harvey’s Legal Agent Benchmark. Note that the finance and authorized outcomes are reported as relative wins, with out absolute scores within the announcement.

Flash Cyber, and why it’s gated

On CyberGymnasium, the usual vulnerability discovery benchmark, Google studies frontier-level efficiency that surpasses each 3.5 Flash Cyber and considerably bigger frontier fashions, although no absolute determine is printed. Because CyberGymnasium is usually C and C++, Google additionally ran an inside benchmark throughout 20 programming languages and studies a discovery success price above 70%.

For patching, CWE-Bench, run by Collinear, places Flash Cyber at 47.2% cross@1 in opposition to a number one frontier mannequin’s 47.8%. Near-parity at materially decrease price is the declare, and Google frames it as sitting on the Pareto frontier moderately than topping the leaderboard.

Chrome Security studies 2.6x extra right patches than one of the best, a lot bigger industrial fashions. Wiz measures 7.5 to 9.7 proportion factors greater recall on its inside penetration testing benchmark at 2.3x to five.2x decrease price. Google’s Cloud Vulnerability Research group discovered a essential foundational vulnerability in below two hours, work that usually takes months.

Google is specific that it prioritized vulnerability fixing over offensive capabilities like exploitation. Flash Cyber ships with a extra permissive set of cyber mitigations, which is exactly why it’s restricted to trusted defenders: authorities authorities, essential infrastructure operators, and software program maintainers who apply by means of Fairwind.

Interactive explainer

Key Takeaways

  • Two variants ship on one shared core, break up by security mitigations moderately than mannequin measurement.
  • 3.8 Flash holds 3.7 Flash pricing at $0.75/$3.75 per 1M tokens by means of December 31, 2026.
  • It trades tokens for accuracy: extra reasoning steps, extra device calls, greater spend per activity.
  • Flash Cyber hits 47.2% cross@1 on CWE-Bench in opposition to a frontier mannequin’s 47.8%, at far decrease price.
  • Cyber entry is gated to vetted defenders, not offered on a worth sheet.

Check out Technical details here. Also, be at liberty to comply with us on Twitter and don’t overlook to hitch our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

Need to associate with us for selling your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar and so on.? Connect with us

The put up Google DeepMind Releases Gemini 3.8 Flash and Gemini 3.8 Flash Cyber: One Core Model, Two Access Envelopes appeared first on MarkTechPost.

Similar Posts