AI
Google Releases Third Gemini Flash Model in Six Weeks, Targeting Agentic Tasks
The company shipped Gemini 3.8 Flash with improved performance on coding and automated reasoning, alongside a restricted-access cybersecurity variant for trusted security partners.
The New Stack
Google has rolled out a fresh iteration of its Gemini Flash lineup, continuing its rapid cadence of model releases. The latest version, Gemini 3.8 Flash, arrives just three weeks after Gemini 3.7 Flash and marks Google's third Flash release within the span of six weeks—a shift in momentum after an extended pause in Gemini Pro updates.
Performance Gains on Coding and Autonomous Systems#
The new Flash 3.8 demonstrates notable improvements over its predecessor, particularly for agentic coding tasks and autonomous computer use applications. On engineering-focused benchmarks like DeepSWE, the model matches Anthropic's Claude Opus 5 and outperforms both OpenAI's GPT-5.6 Sol and Claude Sonnet 5.
However, Google acknowledges that reaching these performance levels comes at a computational cost. The model exercises greater diligence on complex tasks—executing additional reasoning steps and calling tools iteratively—which translates to higher token consumption. Developers can mitigate this by selecting from different reasoning modes to balance performance against cost.
The benchmarks notably exclude Anthropic's recently released Fable 5.1, which demonstrates stronger overall performance on general agentic tasks (55.8% on Terminal-Bench 4.0 compared to Flash 3.8's 19.1%) but commands significantly higher pricing. Flash 3.8 does maintain an edge on Terminal-Bench 2.1, which focuses specifically on coding work.
Google credits its accelerated development cycle to agentic loops—automated evaluation and refinement processes applied during model training itself.
Pricing and Availability#
Flash 3.8 carries an introductory rate of $0.75/$3.75 per million input/output tokens through December 31, 2026, after which it rises to $1.50/$7.50. The model is now available across Google's ecosystem, including Google AI Studio, Antigravity, Android Studio, Gemini Enterprise, and consumer products like the Gemini App and Google Search's AI mode.
Fairwind: A Restricted Cybersecurity Variant#
Alongside the general-purpose Flash model, Google introduced Gemini 3.8 Flash Cyber, a specialized version restricted to approximately 650 "trusted defenders" through Google's newly named Fairwind Program. This approach mirrors Anthropic's strategy of gating high-capability security models.

