Google Launches Gemini 3.8 Flash with Tunable Reasoning and Cyber Variant

Google released Gemini 3.8 Flash on September 2, 2026, marking its third Flash model update in a six-week period. [2][3] The new model features a one-million-token context window, supports up to 64,000 output tokens, and includes tunable thinking levels of low, medium, and high. [3]

Google noted that the model works harder on complex tasks by executing additional reasoning steps and calling tools iteratively, which may result in higher token usage. [3][5] The model's knowledge cutoff extends to March 2026 for some domains, while remaining limited to January 2025 for others. [5]

Google introduced Gemini 3.8 Flash at an introductory rate of $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. [3] Standard pricing of $1.50 per million input tokens and $7.50 per million output tokens will take effect on January 1, 2027. [3]

Alongside the standard model, Google introduced Gemini 3.8 Flash Cyber, a restricted variant for vulnerability detection and automated patching available to trusted defenders via the Gfaiwind (also referred to as Fairwind) programme. [3][4] Google stated that rigorous training in the cybersecurity domain contributed to the coding and reasoning improvements across the shared core of both models. [4]

According to Google, Gemini 3.8 Flash frequently approaches the performance of higher-cost frontier models while operating at a fraction of the cost. [4][5] The model outperforms most larger frontier models in autonomously solving complex end-to-end engineering problems on the long-horizon DeepSWE v1.1 benchmark. [5]

In fields requiring advanced analysis and reporting, the model surpassed Gemini 3.7 Flash and other frontier models on the Vals Finance Agent V2 and Harvey's Legal Agent Benchmark. [5] Gemini 3.8 Flash also scored 54.9 percent on HLE-Verified, reflecting its capacity for multi-step reasoning across STEM, humanities, and professional disciplines. [5]

Google’s aggressive release cadence—delivering three Flash updates in just six weeks—indicates a highly competitive race to dominate the efficiency-focused mid-tier AI market. By offering tunable thinking levels, Gemini 3.8 Flash mirrors the dynamic inference-compute strategies seen in models like OpenAI's o1, allowing developers to balance latency, cost, and reasoning depth for specific agentic workflows. Furthermore, gating the Flash Cyber variant behind the Gfaiwind programme highlights a growing industry trend of restricting frontier cybersecurity capabilities to trusted access channels rather than open APIs. What the sources don't address: Whether the increased token consumption from the model's iterative reasoning loops will negate the cost savings of the introductory pricing for enterprise deployments.