- Published on
Gemini 3.8 Flash Released at Light Speed: Efficient but Lacking in Insight
- Authors
- Name
- DeepHorus
Google officially released its latest AI model, Gemini 3.8 Flash, in the early hours of today. This marks the third Flash series model launched by Google within a mere six weeks, demonstrating its commitment to a "small steps, fast pace" strategy. The new model is positioned to handle more complex long-horizon software engineering, autonomous Agent tasks, and enterprise workflows, indicating Google's intention to have lightweight models take on more flagship-level responsibilities.
According to official benchmarks released by Google, Gemini 3.8 Flash has achieved significant improvements in several key capabilities. In the long-horizon software engineering test DeepSWE v1.1, it scored 71.0%, an increase of nearly 6 percentage points from the previous generation 3.7 Flash's 65.3%, approaching Claude Opus 5's 74.0%. In the multi-domain HLE-Verified test, it slightly surpassed Opus 5 with a score of 54.9% versus 54.4%. Its performance was even more impressive in multimodal tasks such as chart comprehension and video processing.
Despite the performance boost, Gemini 3.8 Flash's pricing remains consistent with the previous 3.7 Flash: $0.75 per million tokens for input and $3.75 for output. This promotional pricing will last until January 1, 2027, after which it will revert to the original price. Google explained that the performance improvement mainly comes from the model "trying harder" to reason and use tools, but this might lead to it consuming more tokens. For efficiency-focused tasks, Google suggests users can lower the reasoning tier or use 3.7 Flash.
Alongside the main model, Google also released Gemini 3.8 Flash Cyber, focused on cybersecurity. This model is designed to help cybersecurity agencies discover and fix vulnerabilities. In internal tests, it achieved a success rate of over 70% in vulnerability discovery tasks covering 20 programming languages. However, this specialized model is currently only available to trusted cybersecurity agencies through the Fairwind Program and is not directly accessible to general users.
However, community feedback and initial third-party tests have questioned the official claims. While users generally acknowledge 3.8 Flash's response speed, in actual programming and complex task generation, the model often shows hasty understanding of requirements and rough details in the results. For example, in a test generating a specific 3D model, although the model quickly outputted code, the final result's details and accuracy fell far short of expectations. This reveals the model's limitations in independently handling complex instructions.
Analysis suggests the discrepancy between official demos and actual user experience might stem from fundamentally different usage patterns. Google's official cases typically run within a carefully designed Agent framework and tool environment, whereas users are more inclined to use one-off natural language instructions. This indicates Gemini 3.8 Flash may perform well as an Agent component but still has a gap in independently supporting a reliable autonomous workflow.
Recently, Koray Kavukcuoglu, the new head of Google DeepMind, publicly stated that the company will not abandon competition in the frontier model arena. However, in practice, after encountering setbacks in developing its flagship model Gemini 3.5 Pro, Google has shifted more resources toward the iteratively faster and lower-cost Flash series. This strategy aims to stabilize users, developers, and the market through Flash series models before the long-term goal of Gemini 4 truly arrives.
Meanwhile, the "frontier" competition in AI also brings new challenges. Other companies' frontier models showcase powerful capabilities but also introduce harder-to-control, higher inference costs, and potential safety and regulatory issues. From this perspective, Google's pragmatic approach focusing on the Flash series might find a sustainable development path in the fierce market competition by better balancing capability, cost, and controllability, beyond just pursuing cutting-edge performance.
