Google launches Gemini 3.8 Flash and Flash Cyber for security

Google DeepMind released Gemini 3.8 Flash and a specialized sibling, Gemini 3.8 Flash Cyber, on September 2, 2026, according to the company's own announcement and reporting by The Verge. The launch lands just weeks after Gemini 3.7 Flash, making 3.8 Flash the third Flash-tier model Google has shipped in six weeks — a pace that underlines how quickly the company is iterating on its cheaper, faster model line.

Same price, but a model that 'works harder'

Gemini 3.8 Flash keeps the introductory pricing Google set for 3.7 Flash: $0.75 per million input tokens and $3.75 per million output tokens, The Verge reports. Google itself has cautioned that the sticker price does not tell the whole story. The company says the new model "works harder" than its predecessor — taking more reasoning steps and calling tools iteratively before it answers. Because Gemini is billed per token rather than per query, a model that reasons longer and calls tools more often can end up costing more in practice, particularly at higher effort settings, even though the price per token has not moved.

Independent benchmarking firm Artificial Analysis offers a different read on the economics. It called Gemini 3.8 Flash "the cheapest we have measured at this intelligence level," estimating it costs roughly 40% less than Gemini 3.7 Flash for equivalent intelligence. That efficiency gain comes with a trade-off pulling the other way: the firm found the model uses about 30% more output tokens per task and takes more turns on agentic evaluations, so cost per unit of intelligence improves even as raw token consumption climbs.

Opus 5 coding quality but at a fraction of the cost and super fast... going to be so awesome for things like making remotion videos

Beating rivals on coding, finance and legal benchmarks

On the DeepSWE v1.1 software engineering benchmark, Gemini 3.8 Flash outperforms both its predecessor and other frontier models, including Claude Fable 5 — which Anthropic itself updated with a lower price this same week, according to Google DeepMind's announcement. Google also reports that the model leads competitors on the Vals Finance Agent V2 benchmark and on Harvey's Legal Agent benchmark, two evaluations built around real professional workflows in finance and law rather than generic question-answering.

On safety, Google DeepMind says Gemini 3.8 Flash carries built-in safeguards against misuse in chemical, biological, radiological and nuclear (CBRN) contexts as well as cyber-offensive use, and scores 54.9% on HLE-Verified, a benchmark of expert-level questions. The company also reports a notable improvement in the model's robustness to prompt-injection attacks, as measured by third-party red-teaming firm Gray Swan.

Flash Cyber and a new program for governments

Alongside the general-purpose model, Google introduced Gemini 3.8 Flash Cyber, a specialized version built for defensive cybersecurity work. It ships together with a new Fairwind Program restricted to governments and what Google calls "trusted partners" — a group it says already counts 650 members, including CrowdStrike and the Center for Internet Security. Members get access to Flash Cyber as well as Google's CodeMender agent, which the company describes as able to "autonomously find and fix vulnerabilities, protecting critical infrastructure, public services, and national security."

• CyberGym: frontier-level performance on autonomous vulnerability discovery, ahead of Flash Cyber 3.5 and larger frontier models
• Internal benchmark across 20 programming languages: success rate above 70%
• CWE-Bench automated patching (run by Collinear): 47.2% pass@1, against 47.8% for a leading frontier model, at markedly lower cost
• Google Chrome security team: 2.6 times more correct fixes than the best commercial models
• Wiz penetration-testing benchmark: 7.5 to 9.7 percentage points more recall, at 2.3 to 5.2 times lower cost
• Google Cloud vulnerability research team: found a critical, fundamental flaw in under two hours, on research that normally takes months

Gemini 3.8 Flash itself is available immediately to Google AI Pro and Ultra subscribers, to developers through the Gemini API, Google AI Studio and Android Studio, and to businesses through Gemini Enterprise, according to Google DeepMind.

What falling agentic-coding costs mean outside the US

For developers and startups outside the United States, the number that matters most is the one Artificial Analysis put on the table: roughly 40% less cost per unit of intelligence than the previous Flash generation. Teams building coding agents, automated patch pipelines or vertical AI tools in markets without deep venture funding can reach frontier-adjacent coding and security capability through the same public Gemini API and Google AI Studio channels used everywhere else — no dedicated enterprise contract required to start experimenting, though access to CodeMender through the Fairwind Program stays limited to governments and vetted partners for now.