Gemini 3.8 Flash and Flash Cyber: Pricing, Benchmarks, and Who Gets Access

Google released Gemini 3.8 Flash and 3.8 Flash Cyber on September 2. Pricing stays at the 3.7 Flash introductory rate of $0.75/$3.75 per million tokens, but it expires December 31, 2026 and doubles. The cyber variant ships only to vetted Fairwind Program organizations.

Gemini Google LLM security API coding

Google released two models on September 2, 2026: Gemini 3.8 Flash and Gemini 3.8 Flash Cyber¹. The launch follows 3.7 Flash from three weeks earlier and marks the third Flash release in six weeks¹.

What stands out is how differently the two are distributed. Gemini 3.8 Flash ships broadly — the Gemini API, the Gemini app, AI Mode in Google Search. Gemini 3.8 Flash Cyber goes only to organizations vetted through the Fairwind Program, which launched the same day¹². Both run on the same foundational intelligence¹.

The price holds, but only through year-end

Gemini 3.8 Flash costs the same as 3.7 Flash: $0.75 per million input tokens and $3.75 per million output tokens¹.

That is an introductory price, and it expires on December 31, 2026¹. A footnote on the announcement page states that $1.50 input and $7.50 output apply starting January 1, 2027¹. The structure is identical to the one used for Gemini 3.7 Flash, which was also half price through year-end and doubles in the new year. If you are budgeting API spend annually, the model name changed but the January 1, 2027 cutover did not.

On performance, Google positions 3.8 Flash as its “most intelligent workhorse model,” claiming significant improvements over 3.7 Flash across software engineering, agentic tasks, and multi-step reasoning in specialized domains¹. The one figure stated in the announcement text is 54.9% on HLE-Verified, which Google describes as evidence of multi-step reasoning across STEM, humanities, and professional fields¹. On DeepSWE v1.1 (long-horizon software engineering), Google says 3.8 Flash outperforms most larger frontier models; on quantitative and professional benchmarks such as Vals Finance Agent V2 and Harvey’s Legal Agent Benchmark, it says the model outperforms 3.7 Flash and other frontier models¹. The announcement text gives no scores for any of them¹.

It is designed to spend more tokens

Google attributes the performance gains less to raw capability than to a design choice: 3.8 Flash “works harder”¹. On complex tasks it executes extra reasoning steps and calls tools iteratively. As a result, the announcement states plainly that the model might use more tokens to maximize performance, especially at higher effort levels¹.

This is the part a price-per-token comparison misses. If the same job consumes more tokens, the bill can rise even when the unit price holds. Google acknowledges this and points to two options for efficiency-first workloads: use lower effort levels, or stay on 3.7 Flash, which remains fully supported¹. Measuring token consumption and accuracy on your own workload before switching is the sensible path.

Choosing who receives the cyber variant

Gemini 3.8 Flash Cyber is built for vulnerability detection and automated patching¹. Google offers the following figures from its own evaluations.

On an internal benchmark spanning complex codebases in 20 programming languages, the model exceeds a 70% success rate at discovering vulnerabilities¹. On CWE-Bench, an external patching benchmark run by Collinear, it reaches 47.2% pass@1 against a leading frontier model’s 47.8% — close enough that Google calls it “on the Pareto frontier” given the significantly lower cost¹. Internally, Google says its Chrome Security team found the model produced 2.6 times more correct patches than the best commercial models that are much larger, and its Cloud Vulnerability Research team found a critical foundational vulnerability in under two hours, work that usually takes months¹. Security company Wiz reported 7.5–9.7% higher recall on its internal penetration testing benchmark at 2.3–5.2x lower cost than other leading frontier models¹.

Google states it has invested in vulnerability fixing from the start and prioritized it over offensive capabilities like exploitation¹. Even so, this model is distributed nothing like 3.8 Flash.

Fairwind as the gate

The Fairwind Program launched the same day as a limited-access program for governments and trusted partners². It targets a trusted group of Google Cloud customers, government agencies, and cybersecurity partners, and pairs Gemini 3.8 Flash Cyber with the CodeMender harness². Google says defenders can generate verified, deployment-ready patches in minutes within their own secure cloud environment, instead of spending weeks on manual fixes².

Staged access covers three categories: governments and national cyber authorities; critical infrastructure operators across healthcare, telecommunications, energy, and financial networks; and core technology platforms securing widely used software foundations². Participating organizations agree to operational standards, including limiting access to employees on internal cybersecurity, incident response, or penetration testing teams, and deploying protections such as multi-factor authentication². Google says there are more than 650 participating partners globally².

Google frames the reasoning as giving defenders an “adaptation window” to harden their systems before bad actors can exploit new capabilities². The safety treatment differs between the two models as well: 3.8 Flash ships with safeguards against misuse in CBRN (chemical, biological, radiological, and nuclear) and cyber offense domains, while 3.8 Flash Cyber ships with a more permissive set of mitigations for cybersecurity¹. That is precisely why it is restricted to trusted defenders who need a more comprehensive set of cyber capabilities¹.

Google Cloud customers outside Fairwind can still use CodeMender with publicly available models hosted on the Gemini Enterprise Agent Platform to secure their code². What is gated is the Cyber model, not the patching machinery itself.

The Flash Cyber line, and another announcement the same week

This is not the first cyber-focused Flash model. The line began with the Gemini 3.6 Flash, 3.5 Flash-Lite, and Flash Cyber release on July 21, and even then Flash Cyber was not generally available — it went to governments and trusted partners through a pilot program built around CodeMender. What is new is that the access arrangement now has a name and stated participation terms in the Fairwind Program. The Flash line also moves quickly: agentic video understanding shipped for 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite just a day earlier, on September 1.

In broader context, “release a highly capable cyber model only to vetted recipients” appeared twice this week. On September 1, OpenAI said its upcoming Astra model had reached the Critical cybersecurity threshold under its Preparedness Framework, and that access to advanced cyber capabilities would initially go to a small group after release. The two companies reached that conclusion through different reasoning and different processes, but the conclusion — do not hand capable models to everyone — is the same.

For teams running security operations, the practical difference sits in the access terms. Models like the Cyber variant are no longer something you obtain by reading a price sheet and signing up. Applying, passing review, and deciding which internal teams get access has become the first piece of work in evaluating adoption.