Google Ships Third Gemini Flash Model in Six Weeks

Google released Gemini 3.8 Flash on September 2, 2026, just three weeks after Gemini 3.7 Flash and roughly six weeks after Gemini 3.6 Flash. That's three Flash-tier model updates since July 21, an unusually tight release cadence even by the standards of a company that ships models frequently.

The new model pushes harder on reasoning and tool use. Google says Gemini 3.8 Flash takes more intermediate reasoning steps on complex tasks and calls external tools iteratively rather than in a single pass, a shift aimed at agentic workflows where a model needs to check its own work mid-task.

That extra reasoning shows up in benchmarks. Gemini 3.8 Flash scored around 71% on the DeepSWE v1.1 software engineering benchmark, up from 65.3% for its predecessor, and ahead of Anthropic's Fable 5 (Claude) on the same test.

Pricing stays flat at $0.75 per million input tokens and $3.75 per million output tokens, unchanged from Gemini 3.7 Flash. But flat per-token pricing doesn't mean flat bills: Google estimates the model's heavier reasoning generates roughly 30% more output tokens per task, meaning real-world costs could climb even though the sticker price didn't move.

The model handles a one-million-token context window across text, image, audio, and video input, with output capped at 65,536 tokens. It's rolling out to consumers on Google AI Pro or Ultra plans, alongside developers and enterprise customers through the API.

A Separate, Restricted Release

Alongside the consumer-facing launch, Google introduced Gemini 3.8 Flash Cyber and a new Fairwind Program, both restricted to governments and vetted partners rather than the general public. The program already counts 650 members, including CrowdStrike and the Center for Internet Security.

Fairwind gives approved members access to CodeMender, Google's agent for autonomously finding and patching software vulnerabilities, pitched specifically at protecting critical infrastructure and national security systems. It's a narrower, security-focused track running parallel to the mainstream Flash release.

Google also built in explicit safeguards against misuse in chemical, biological, radiological, and nuclear (CBRN) contexts as well as cyber-offense applications, reflecting the same dual-use concerns that shaped the decision to gate CodeMender access rather than release it broadly.

The pace of these Flash releases, three distinct versions in six weeks, suggests Google is treating rapid iteration as a competitive lever against rivals like Anthropic and OpenAI, even as each incremental jump in capability now carries a corresponding governance question for restricted, higher-risk deployments like Fairwind.