Google released two new Gemini models on September 2, and the one that matters most for practitioners is the pricing sheet. Gemini 3.8 Flash ships at $0.75 per million input tokens and $3.75 per million output tokens, identical to the 3.7 Flash that preceded it three weeks earlier. The companion model, Gemini 3.8 Flash Cyber, focuses on vulnerability discovery and automated patching, and it is not open for general access. It reaches trusted defenders through a new program Google is calling Fairwind.
The announcement came as a joint blog post from Tulsee Doshi, senior director of product management, and Raluca Ada Popa, the Gemini security lead at Google DeepMind, in a post introducing Gemini 3.8 and Flash Cyber (Google blog). They framed the release as the third Flash variant in just six weeks, a cadence that reads less like methodical product development and more like a pricing war being fought on capability gains.
What changed in 3.8 Flash
The headline improvement is agentic coding. Google says 3.8 Flash outperforms "most larger frontier models" on DeepSWE v1.1, a long-horizon software engineering benchmark that asks a model to solve complex engineering tasks end to end without hand-holding. On a separate reasoning measure, the model scores 54.9% on HLE-Verified, which spans STEM, humanities, and professional problems requiring multi-step chains.
Google's explanation for the jump is a behavioral one rather than an architectural one. The model "works harder" on hard tasks, running more reasoning steps and calling tools iteratively, spending more tokens specifically to maximize performance. For a data scientist, this is the classic inference-time compute tradeoff: you are buying accuracy by burning tokens. The question for any team benchmarking this model is whether the accuracy gain per extra token clears the hurdle your task sets.
What Flash Cyber actually does
Flash Cyber is the more interesting artifact, and it is where Google points to real engineering outcomes instead of benchmark tables. The model is measured against CyberGym, the standard industry test for finding vulnerabilities, and against CWE-Bench, an external patching benchmark run by Collinear. On the latter, it lands at a pass@1 of 47.2% versus a leading frontier model at 47.8%, but Google notes that figure comes at a significantly lower cost.
An internal test spanning 20 programming languages pushed success above 70%, and the company cites concrete production use. Chrome's security team reportedly found that Flash Cyber produced 2.6 times more correct patches to vulnerabilities in Chrome than the best commercial models, even though those models are much larger. Wiz, which Google partnered with, recorded 7.5 to 9.7% higher recall on its penetration testing benchmark at 2.3 to 5.2 times lower cost.
The ordering of claims is telling. Google lists vulnerability discovery and patching well before it lists offensive capabilities like exploitation, a design choice that separates this model from the OpenAI Astra debate from earlier this week, where opaque recurrence and auditability raised real eyebrows.
Our read
From an engineering standpoint, the two numbers worth carrying away are the token prices and the 2.6x patch count. The pricing is not a discount, but it is a refusal to raise price on a capability upgrade. When a model gets meaningfully better at a task while the price-per-token holds, the economics shift entirely in favor of routing real work through it instead of reserving it for internal tooling.
The 2.6x correct-patch rate against larger commercial models is the more consequential figure, and it deserves the skepticism it will get. "Correct patches" is a narrow label. A patch can be correct on a benchmark and still misfire against a codebase with unusual dependency chains, odd linting rules, or partial mocks. The internal Chrome example is useful precisely because it is production-adjacent, but sample size and selection criteria are not disclosed. Treat 47.2% versus 47.8% as roughly equal performance with a clear cost asymmetry, and treat the 7.5 to 9.7% recall gains as the signal worth validating on your own workloads before trusting them.
From a data science angle, the "works harder" design matters more than the benchmark deltas. A model that spends more tokens on hard tasks is effectively doing self-adaptive inference-time compute. That behavior is easy to benchmark well and painful to control in production, because you no longer know in advance which requests will trigger the expensive mode. Teams should budget for variable cost, not a flat per-call price, when deploying this kind of model into an agentic pipeline.
How it compares and where it lands
This release sits directly alongside the frontier moves from OpenAI. Astra cleared a Critical cybersecurity threshold before launch, which we covered earlier this week, and its release stayed gated precisely because of the capabilities Flash Cyber also demonstrates. Google is simply choosing to ship those capabilities at Flash pricing and channel Flash Cyber through a vetted access program rather than a gated frontier release.
For developers, the practical move is straightforward. Route long-horizon coding and agentic tool use through 3.8 Flash, sized for variable token spend. Reserve Flash Cyber for vulnerability discovery and patching work where the cost per correct patch already clears your security budget. Both are documented on Google's blog and the official model card, where the full benchmark breakdown lives.