Google has introduced Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, two new AI models aimed at improving efficiency for enterprise AI agents. The company says Gemini 3.6 Flash generates 17% fewer output tokens than the previous version, helping reduce operational costs and latency for organizations running large-scale automated workflows.
According to Google, Gemini 3.6 Flash achieved a 49% success rate on the DeepSWE benchmark, compared with 37% for Gemini 3.5 Flash. On the MLE Bench benchmark, performance increased from 49.7% to 63.9%, while its GDPval-AA v2 score improved from 1349 to 1421.
Several companies, including Figma, Harvey, and Hebbia, have integrated the model into their workflows. These organizations use it for tasks such as software prototyping, document analysis, financial filing reviews, and report generation.
Google also announced a built-in computer-use tool for Gemini models, allowing AI systems to interact directly with operating systems without requiring custom intermediary software. The company reported an OSWorld-Verified score of 83.0%, up from 78.4%.
For high-volume tasks, Gemini 3.5 Flash-Lite offers lower-cost processing with pricing of $0.30 per million input tokens and $2.50 per million output tokens. Google states that the model can generate approximately 350 output tokens per second, making it suitable for document processing and agentic search workloads.
In cybersecurity, Google introduced Gemini 3.5 Flash Cyber, a restricted model designed for vulnerability validation and remediation. Access is currently limited to governments and vetted partners through a pilot program. The model is used within Google’s CodeMender security agent, where multiple AI instances review vulnerabilities and produce remediation recommendations for human approval.
Google also confirmed that Gemini 3.5 Pro remains in partner testing and that development of the future Gemini 4 architecture is already underway.
