Gemini 3.8 Flash is a foundational model released in September 2026 as the latest iteration of the Gemini 3 series. It serves as an incremental enhancement to Gemini 3.7 Flash, specifically engineered to optimize performance in software engineering tasks and agentic knowledge workflows while maintaining the capacity for customizable effort levels to balance quality, latency, and cost.

Core Architecture and Specifications

The model utilizes a multimodal architecture capable of processing text, images, audio, and video inputs. It features a 1M token context window and a 64K token output capacity. As an iterative update based on the Gemini 3.7 Flash architecture, training datasets, and hardware/software infrastructure, Gemini 3.8 Flash inherits its predecessor's foundational design while integrating refined capabilities. The model's knowledge cutoff is split-domain: some domains include information up to March 2026, while others remain capped at January 2025.

Evaluation Methodology

The evaluation approach for Gemini 3.8 Flash spans diverse domains, including scientific reasoning, long-context comprehension, computer use, and multimodal processing. Performance metrics are derived from automated evaluations, with the caveat that these results are not directly comparable to previous versions due to refinements in the evaluation methodologies themselves.

Safety and ethical performance were measured through specific automated metrics:

• Text-to-Text Safety: Measured with a performance delta of $-0.4\text{pp}$ (where lower is better).
• Multilingual Safety: Exhibited a regression of $+5.4\text{pp}$ (where lower is better).
• Tone: Evaluated for objective tone on sensitive topics, showing a $+0.2\text{pp}$ improvement.
• Unjustified Refusals: Measured by the model’s ability to handle borderline prompts, showing a $+1.1\text{pp}$ increase in sensitivity.

Frontier Safety and Governance

Gemini 3.8 Flash was assessed under the April 2026 Frontier Safety Framework. Because the model demonstrates no material increases in capabilities or new functionalities compared to Gemini 3.7 Flash, it is considered unlikely to reach any Tracked or Critical Capability Levels (T/CCLs).

The safety assessment protocol involves both automated content safety evaluations and external human red teaming. For child safety, the model met established launch thresholds. Manual reviews of performance fluctuations in automated tests concluded that identified regressions were predominantly false positives rather than indicators of egregious or dangerous model behavior. The model is distributed across Google’s ecosystem, including the Gemini app, AI Studio, Gemini API, and the Enterprise Agent Platform.