What is Google’s reportedly imminent Gemini 3.8 Flash coding AI model, internally known as “Skimaki,” when might it be released, how did Google engineers compare it with Anthropic’s Claude Opus using the internal Jetski coding tool, why is it designed as a faster and cheaper Flas

Gemini 3.8 Flash, reportedly codenamed “Skimaki,” was described as a coding-focused, efficient Gemini model intended to close Google’s perceived gap with OpenAI and Anthropic. The report said it could launch as early as Wednesday, September 2; a Google developer page indexed that day also described Gemini 3.8 Flash as generally available, so the “imminent” framing may already be outdated. 13

• What it is: A “Flash” model is meant to prioritize speed and lower inference cost rather than be Google’s largest frontier-scale system. The reported target is long-horizon software engineering and agentic workflows—areas in which models must repeatedly inspect code, use tools, make changes, and validate results. 13

• Internal comparison with Claude: According to employees cited by the Wall Street Journal, Google engineers ran head-to-head coding comparisons in Jetski, an internal coding tool, and some preferred 3.8 Flash to Anthropic’s Claude Opus. That is an internal, qualitative evaluation—not an independently reproducible public benchmark or proof of general superiority. 1

• Why a Flash model: The reported strategy is to get strong coding/agent performance at lower latency and cost, making it more practical for high-volume product use. The claim that this avoids the heavier scrutiny associated with a frontier-scale model is a strategic inference from its smaller, efficiency-oriented positioning; the available reporting does not establish a specific government-vetting threshold or confirm that regulation was the decisive design constraint. 1

• Training approach: Google reportedly increased resources devoted to reinforcement learning—the later training stage in which models learn skills through trial, feedback, and reward—to improve task execution. This helps explain the quick succession from Gemini 3.7 Flash to the reported 3.8 model: the emphasis is on converting base-model capability into more reliable coding and tool-use behavior, rather than only scaling pretraining. 1

• Leadership and organizational context: In August, Demis Hassabis moved from running Google DeepMind day to day to become its chair and Alphabet’s chief scientist; the change put more operational authority with other executives while preserving Hassabis’s strategic/scientific role. 29 Reuters also reported that Sergey Brin urged key AI staff to go “all in” on Gemini, while Google has positioned AI agents as central to commercializing its AI efforts. 14

• Competitive purpose: The model appears aimed at a practical weakness: Google has been viewed as behind OpenAI and Anthropic in coding, particularly as OpenAI shifted its focus toward agentic coding. A fast, cheaper model that can power coding agents at scale would be more deployable across Google products and cloud customers than a costly flagship alone. 14

• What remains unproven: Strong internal Jetski results may not carry over to public benchmarks, diverse real-world repositories, reliability over long autonomous runs, security, or price-performance after release. The evidence for the Opus comparison comes from unnamed employees, and public performance depends on the released model, its tool environment, evaluation methodology, and how quickly OpenAI and Anthropic respond. Insufficient evidence supports any definitive claim that 3.8 Flash has surpassed them overall. 1