Google launches Gemini 3.8 Flash with more expensive tasks

Google adds Gemini 3.8 Flash to its series of economic models, three weeks after the release of version 3.7. Internal and independent measures credit it with better coding results and improved robustness, with the cost per token remaining unchanged. However, the cost per task has increased, and the Cyber variant remains reserved for verified entities.

Independent Measures: Cost per Task +40% and 2.5 Minutes on Average

Artificial Analysis assigns Gemini 3.8 Flash an Intelligence Index of 59, compared to 56 for 3.7 Flash. GPT-5.6 Sol in high reasoning mode and Grok 4.6 in medium mode also achieve a score of 59. According to Artificial Analysis, this gain primarily comes from better performance on agentic tasks such as tool usage and coding. In terms of cost, 3.8 Flash reaches the Pareto frontier at $0.58 per task, making it the cheapest model at this level of intelligence, but this amount is about 40% higher than the $0.40 measured for 3.7 Flash, while the price per token has not changed. When it comes to tasks requiring advanced reasoning, 3.8 Flash produces nearly 300 tokens per second in output, and the average time for each task is 2.5 minutes. This duration is slightly lower than that of GPT-5.6 Luna (2.6 minutes) and GPT-5.6 Terra (3.3 minutes), but remains higher than Claude Fable 5.1 (2.1 minutes) and 3.7 Flash (2.2 minutes). By lowering the reasoning level, the average duration drops to about 48 seconds.

Cybersecurity: Access via Fairwind and 86.2% on CyberGym

Gemini 3.8 Flash Cyber is not available to the public, and access is granted through the Fairwind Program, aimed at government agencies, critical infrastructure operators, and software maintenance managers. This version employs less strict security rules than the standard version, as it targets applications in defensive cybersecurity. For the CyberGym benchmark dedicated to vulnerability detection in C/C++ code, Google reports a result of 86.2%, surpassing 3.5 Flash Cyber (77.5%), GPT-5.6 Sol (83.6%), and GPT-5.5-Cyber (85.6%). On the external CWE-Bench assessment for automated patching, the model achieves 47.2% Pass@1, close to the benchmark level of 47.8%, with a much lower cost. Regarding robustness against prompt injection attacks, Gemini 3.8 Flash shows a 5.5% attack success rate on Gray Swan IPI. DeepSeek V4 Pro, Kimi K3, and Grok 4.6 are measured at 60.1%, 52.7%, and 51.8%. Claude Opus 5 drops to 4.8%, and its enhanced security options further lower this score within the Claude ecosystem.

Unchanged Price per Token but More Tokens Consumed

Gemini 3.8 Flash maintains a launch price of $0.75 per million input tokens and $3.75 per million output tokens, identical to 3.7 Flash. Google plans to transition to regular pricing of $1.50 and $7.50 starting January 2027. For comparison, Claude Opus 5 is charged at $5.00 for input and $25.00 for output, while GPT-5.6 Sol is priced at $4.00 and $20.00. Even after the end of promotional rates, the cost per token of 3.8 Flash would remain lower than that of OpenAI and Anthropic's flagship models. Google explains that the performance gains of 3.8 Flash are partly due to additional reasoning steps and iterative tool calls, which increase token consumption and may mitigate the per-token pricing advantage. For workloads where computational efficiency is paramount, Google recommends lowering the reasoning level or opting for 3.7 Flash, which is still supported.

Positioning: Accelerated Pace, Rising Benchmarks, 3D Demonstration

Three weeks after 3.7 Flash, Google introduces Gemini 3.8 Flash in two versions, general and Cyber, as part of a series of three Flash launches in six weeks, while Gemini 3.5 Pro and Gemini 4 are not available. Koray Kavukcuoglu, head of DeepMind, indicates that Google is also aiming for excellence in raw capacity. On DeepSWE v1.1, Google announces 73.7% for 3.8 Flash, close to Claude Opus 5 (74.0%), above GPT-5.6 Sol (72.7%), and an improvement over 3.7 Flash (65.3%), while Claude Sonnet 5 is measured at 53.8%. According to Google, 3.8 Flash achieves better results than Opus 5 and GPT-5.6 Sol on many benchmarks, while being significantly less expensive, emphasizing that benchmark performance does not always reflect real-world usage. Google also highlights advancements in 3D generation, illustrated by the creation of a game from a single prompt in Antigravity, with textures generated by the Nano Banana image model.

Access for Developers, Businesses, and the General Public

Developers can use Gemini 3.8 Flash via Google AI Studio, Google Antigravity, and Android Studio. Businesses access it through Gemini Enterprise. For individuals, the model is available in the Gemini app, in Google Search's AI mode, and in Google Sheets for paying subscribers.