News
Gemini 3.8 Flash ships as Google’s agent workhorse; Flash Cyber stays gated
Google GA’d Gemini 3.8 Flash for long-horizon coding agents and gated 3.8 Flash Cyber to Fairwind partners for vuln find-and-fix.
Searcher → Analyst → Writer → Editor · subagentic-20260902-092013
Google posted two Gemini 3.8 models on September 2, 2026. Only one of them is the coding-and-agent default you can actually put in a loop today.
Gemini 3.8 Flash is the generally available workhorse: long-horizon software engineering, managed agents, and multi-step reasoning at Flash speed and Flash price. Gemini 3.8 Flash Cyber is a separate defender model. It is not a drop-in API twin. Trusted partners get it through Google’s new Fairwind Program, paired with CodeMender for finding and patching vulnerabilities. Treat them as a shared core with two doors, not one interchangeable release.
Google DeepMind framed the split the same day: 3.8 Flash for scaling AI agents, 3.8 Flash Cyber for securing code. On The Keyword, Tulsee Doshi and Raluca Ada Popa called 3.8 Google’s best reasoning and coding model yet, at the same speed and low cost as 3.7 Flash — the previous Flash drop from three weeks earlier, and the third Flash release in six weeks.
Flash is the agent default
Google positions 3.8 Flash as its most intelligent Flash workhorse, with claimed gains over 3.7 Flash across software engineering, agentic tasks, and specialized multi-step reasoning. On DeepSWE v1.1, which the company uses as a long-horizon software-engineering benchmark, it says 3.8 Flash outperforms most larger frontier models at a fraction of the cost. It also claims wins over 3.7 Flash and other frontier models on Vals Finance Agent V2 and Harvey’s Legal Agent Benchmark, plus 54.9% on HLE-Verified — Google’s marker for multi-step reasoning across STEM, humanities, and professional fields.
Those numbers are vendor-reported. The design story underneath them is more useful for practitioners: Google says 3.8 Flash works harder. On complex tasks it takes extra reasoning steps and calls tools iteratively, and it may spend more tokens at higher effort levels. If compute is the constraint, the company points developers at lower effort settings — or at 3.7 Flash, which remains fully supported for efficiency-first workloads.
The Gemini API model page lists the id as gemini-3.8-flash, last updated September 2, 2026. Limits are 1,048,576 input tokens and 65,536 output tokens. Inputs include text, image, video, audio, and PDF; output is text. Thinking is supported at low, medium, and high (minimal returns an error). Computer use is supported in preview. Caching, code execution, file search, function calling, search grounding, Maps grounding, structured outputs, URL context, batch, flex inference, and priority inference are on. Audio generation, image generation, and the Live API are not.
Introductory API pricing matches 3.7 Flash: $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. On January 1, 2027 that becomes $1.50 / $7.50.
Availability is broad on the Flash side. Developers can build in Google Antigravity, the Gemini API via Google AI Studio, Android Studio, and Stitch. Enterprises get it in Gemini Enterprise. Google AI Pro and Ultra subscribers get it in the Gemini app, AI Mode in Search, and Gemini in Google Sheets. Google’s launch demos lean hard on Antigravity: a looping-instruction 3D castle game, a playable DOS-style Google Maps from a single prompt, a USGS-backed topographic explorer, and a Hardware Anatomy teardown visualizer in AI Studio.
That is the product to default to for everyday agent loops. It is also the one with the tighter safety envelope: Google says 3.8 Flash ships with safeguards against misuse in CBRN and cyber offense, aligned with its Frontier Safety Framework, plus a claimed leap in prompt-injection robustness on Gray Swan’s measure.
Flash Cyber is a different product
Gemini 3.8 Flash Cyber is not general availability. Google calls it its most capable cybersecurity model, with frontier-level vulnerability detection and automated patching, at Flash speed and cost — and it is available only to trusted defenders through Fairwind.
On CyberGym, Google says Flash Cyber shows frontier-level autonomous vulnerability discovery and surpasses both 3.5 Flash Cyber and significantly larger frontier models. An internal discovery benchmark spanning 20 programming languages is reported at a success rate exceeding 70%. On Collinear’s CWE-Bench for patching, Google cites a pass@1 of 47.2% against a leading frontier model at 47.8%, at significantly lower cost. Those CyberGym and CWE-Bench figures remain vendor-reported.
Google also cites internal use: Chrome Security found 3.8 Flash Cyber produced 2.6 times more correct patches than the best commercial models it compared, which were much larger; Wiz reported 7.5–9.7% higher recall on an internal penetration-testing benchmark at 2.3–5.2x lower cost; Google’s Cloud Vulnerability Research team says it found a critical foundational vulnerability in less than two hours with the model.
The gating is explicit. Flash Cyber ships with a more permissive set of cybersecurity mitigations than 3.8 Flash, so it is limited to trusted defenders who need a more comprehensive cyber capability set. Google says it invested in vulnerability fixing from the start and prioritized it over offensive capabilities such as exploitation.
Fairwind, launched the same day, is the access path. Four Flynn describes it as a limited program for governments and trusted partners. The first offering pairs Gemini 3.8 Flash Cyber with the CodeMender harness so defenders can find, verify, and fix vulnerabilities at agentic scale, generating verified patches inside an organization’s secure cloud environment. Initial access is staged to governments and national cyber authorities, critical infrastructure operators (healthcare, telecommunications, energy, financial networks), and core technology platforms. Participating organizations agree to operational limits: access restricted to internal cybersecurity, incident response, or penetration-testing teams, plus protections such as multi-factor authentication. Google says more than 650 partners are participating.
Everyone else is not locked out of CodeMender entirely. Google notes that any Google Cloud customer can use CodeMender with publicly available models on the Gemini Enterprise Agent Platform. That is not the same as getting 3.8 Flash Cyber.
What to change in the stack
If you run coding agents in Antigravity, the Gemini API, or Android Studio, 3.8 Flash is the new default to evaluate — especially for long-horizon tasks where Google says diligence and extra tool calls are the point. Watch token spend at high thinking levels; keep 3.7 Flash in the mix where latency and cost dominate. Do not assume Flash Cyber is sitting behind the same model id. It is a Fairwind-only defender stack, not a second flavor of the public Flash endpoint.
If you ship agents, put gemini-3.8-flash in a long-horizon coding loop in Google AI Studio or Antigravity and compare quality and token use against 3.7 Flash at a lower thinking level. If you actually need the cyber variant, read the Fairwind Program post — that is a separate access path, not an interchangeable API drop.